xref: /openbmc/linux/block/bfq-iosched.c (revision eca20409)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Budget Fair Queueing (BFQ) I/O scheduler.
4  *
5  * Based on ideas and code from CFQ:
6  * Copyright (C) 2003 Jens Axboe <axboe@kernel.dk>
7  *
8  * Copyright (C) 2008 Fabio Checconi <fabio@gandalf.sssup.it>
9  *		      Paolo Valente <paolo.valente@unimore.it>
10  *
11  * Copyright (C) 2010 Paolo Valente <paolo.valente@unimore.it>
12  *                    Arianna Avanzini <avanzini@google.com>
13  *
14  * Copyright (C) 2017 Paolo Valente <paolo.valente@linaro.org>
15  *
16  * BFQ is a proportional-share I/O scheduler, with some extra
17  * low-latency capabilities. BFQ also supports full hierarchical
18  * scheduling through cgroups. Next paragraphs provide an introduction
19  * on BFQ inner workings. Details on BFQ benefits, usage and
20  * limitations can be found in Documentation/block/bfq-iosched.rst.
21  *
22  * BFQ is a proportional-share storage-I/O scheduling algorithm based
23  * on the slice-by-slice service scheme of CFQ. But BFQ assigns
24  * budgets, measured in number of sectors, to processes instead of
25  * time slices. The device is not granted to the in-service process
26  * for a given time slice, but until it has exhausted its assigned
27  * budget. This change from the time to the service domain enables BFQ
28  * to distribute the device throughput among processes as desired,
29  * without any distortion due to throughput fluctuations, or to device
30  * internal queueing. BFQ uses an ad hoc internal scheduler, called
31  * B-WF2Q+, to schedule processes according to their budgets. More
32  * precisely, BFQ schedules queues associated with processes. Each
33  * process/queue is assigned a user-configurable weight, and B-WF2Q+
34  * guarantees that each queue receives a fraction of the throughput
35  * proportional to its weight. Thanks to the accurate policy of
36  * B-WF2Q+, BFQ can afford to assign high budgets to I/O-bound
37  * processes issuing sequential requests (to boost the throughput),
38  * and yet guarantee a low latency to interactive and soft real-time
39  * applications.
40  *
41  * In particular, to provide these low-latency guarantees, BFQ
42  * explicitly privileges the I/O of two classes of time-sensitive
43  * applications: interactive and soft real-time. In more detail, BFQ
44  * behaves this way if the low_latency parameter is set (default
45  * configuration). This feature enables BFQ to provide applications in
46  * these classes with a very low latency.
47  *
48  * To implement this feature, BFQ constantly tries to detect whether
49  * the I/O requests in a bfq_queue come from an interactive or a soft
50  * real-time application. For brevity, in these cases, the queue is
51  * said to be interactive or soft real-time. In both cases, BFQ
52  * privileges the service of the queue, over that of non-interactive
53  * and non-soft-real-time queues. This privileging is performed,
54  * mainly, by raising the weight of the queue. So, for brevity, we
55  * call just weight-raising periods the time periods during which a
56  * queue is privileged, because deemed interactive or soft real-time.
57  *
58  * The detection of soft real-time queues/applications is described in
59  * detail in the comments on the function
60  * bfq_bfqq_softrt_next_start. On the other hand, the detection of an
61  * interactive queue works as follows: a queue is deemed interactive
62  * if it is constantly non empty only for a limited time interval,
63  * after which it does become empty. The queue may be deemed
64  * interactive again (for a limited time), if it restarts being
65  * constantly non empty, provided that this happens only after the
66  * queue has remained empty for a given minimum idle time.
67  *
68  * By default, BFQ computes automatically the above maximum time
69  * interval, i.e., the time interval after which a constantly
70  * non-empty queue stops being deemed interactive. Since a queue is
71  * weight-raised while it is deemed interactive, this maximum time
72  * interval happens to coincide with the (maximum) duration of the
73  * weight-raising for interactive queues.
74  *
75  * Finally, BFQ also features additional heuristics for
76  * preserving both a low latency and a high throughput on NCQ-capable,
77  * rotational or flash-based devices, and to get the job done quickly
78  * for applications consisting in many I/O-bound processes.
79  *
80  * NOTE: if the main or only goal, with a given device, is to achieve
81  * the maximum-possible throughput at all times, then do switch off
82  * all low-latency heuristics for that device, by setting low_latency
83  * to 0.
84  *
85  * BFQ is described in [1], where also a reference to the initial,
86  * more theoretical paper on BFQ can be found. The interested reader
87  * can find in the latter paper full details on the main algorithm, as
88  * well as formulas of the guarantees and formal proofs of all the
89  * properties.  With respect to the version of BFQ presented in these
90  * papers, this implementation adds a few more heuristics, such as the
91  * ones that guarantee a low latency to interactive and soft real-time
92  * applications, and a hierarchical extension based on H-WF2Q+.
93  *
94  * B-WF2Q+ is based on WF2Q+, which is described in [2], together with
95  * H-WF2Q+, while the augmented tree used here to implement B-WF2Q+
96  * with O(log N) complexity derives from the one introduced with EEVDF
97  * in [3].
98  *
99  * [1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
100  *     Scheduler", Proceedings of the First Workshop on Mobile System
101  *     Technologies (MST-2015), May 2015.
102  *     http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf
103  *
104  * [2] Jon C.R. Bennett and H. Zhang, "Hierarchical Packet Fair Queueing
105  *     Algorithms", IEEE/ACM Transactions on Networking, 5(5):675-689,
106  *     Oct 1997.
107  *
108  * http://www.cs.cmu.edu/~hzhang/papers/TON-97-Oct.ps.gz
109  *
110  * [3] I. Stoica and H. Abdel-Wahab, "Earliest Eligible Virtual Deadline
111  *     First: A Flexible and Accurate Mechanism for Proportional Share
112  *     Resource Allocation", technical report.
113  *
114  * http://www.cs.berkeley.edu/~istoica/papers/eevdf-tr-95.pdf
115  */
116 #include <linux/module.h>
117 #include <linux/slab.h>
118 #include <linux/blkdev.h>
119 #include <linux/cgroup.h>
120 #include <linux/ktime.h>
121 #include <linux/rbtree.h>
122 #include <linux/ioprio.h>
123 #include <linux/sbitmap.h>
124 #include <linux/delay.h>
125 #include <linux/backing-dev.h>
126 
127 #include <trace/events/block.h>
128 
129 #include "elevator.h"
130 #include "blk.h"
131 #include "blk-mq.h"
132 #include "blk-mq-sched.h"
133 #include "bfq-iosched.h"
134 #include "blk-wbt.h"
135 
136 #define BFQ_BFQQ_FNS(name)						\
137 void bfq_mark_bfqq_##name(struct bfq_queue *bfqq)			\
138 {									\
139 	__set_bit(BFQQF_##name, &(bfqq)->flags);			\
140 }									\
141 void bfq_clear_bfqq_##name(struct bfq_queue *bfqq)			\
142 {									\
143 	__clear_bit(BFQQF_##name, &(bfqq)->flags);		\
144 }									\
145 int bfq_bfqq_##name(const struct bfq_queue *bfqq)			\
146 {									\
147 	return test_bit(BFQQF_##name, &(bfqq)->flags);		\
148 }
149 
150 BFQ_BFQQ_FNS(just_created);
151 BFQ_BFQQ_FNS(busy);
152 BFQ_BFQQ_FNS(wait_request);
153 BFQ_BFQQ_FNS(non_blocking_wait_rq);
154 BFQ_BFQQ_FNS(fifo_expire);
155 BFQ_BFQQ_FNS(has_short_ttime);
156 BFQ_BFQQ_FNS(sync);
157 BFQ_BFQQ_FNS(IO_bound);
158 BFQ_BFQQ_FNS(in_large_burst);
159 BFQ_BFQQ_FNS(coop);
160 BFQ_BFQQ_FNS(split_coop);
161 BFQ_BFQQ_FNS(softrt_update);
162 #undef BFQ_BFQQ_FNS						\
163 
164 /* Expiration time of async (0) and sync (1) requests, in ns. */
165 static const u64 bfq_fifo_expire[2] = { NSEC_PER_SEC / 4, NSEC_PER_SEC / 8 };
166 
167 /* Maximum backwards seek (magic number lifted from CFQ), in KiB. */
168 static const int bfq_back_max = 16 * 1024;
169 
170 /* Penalty of a backwards seek, in number of sectors. */
171 static const int bfq_back_penalty = 2;
172 
173 /* Idling period duration, in ns. */
174 static u64 bfq_slice_idle = NSEC_PER_SEC / 125;
175 
176 /* Minimum number of assigned budgets for which stats are safe to compute. */
177 static const int bfq_stats_min_budgets = 194;
178 
179 /* Default maximum budget values, in sectors and number of requests. */
180 static const int bfq_default_max_budget = 16 * 1024;
181 
182 /*
183  * When a sync request is dispatched, the queue that contains that
184  * request, and all the ancestor entities of that queue, are charged
185  * with the number of sectors of the request. In contrast, if the
186  * request is async, then the queue and its ancestor entities are
187  * charged with the number of sectors of the request, multiplied by
188  * the factor below. This throttles the bandwidth for async I/O,
189  * w.r.t. to sync I/O, and it is done to counter the tendency of async
190  * writes to steal I/O throughput to reads.
191  *
192  * The current value of this parameter is the result of a tuning with
193  * several hardware and software configurations. We tried to find the
194  * lowest value for which writes do not cause noticeable problems to
195  * reads. In fact, the lower this parameter, the stabler I/O control,
196  * in the following respect.  The lower this parameter is, the less
197  * the bandwidth enjoyed by a group decreases
198  * - when the group does writes, w.r.t. to when it does reads;
199  * - when other groups do reads, w.r.t. to when they do writes.
200  */
201 static const int bfq_async_charge_factor = 3;
202 
203 /* Default timeout values, in jiffies, approximating CFQ defaults. */
204 const int bfq_timeout = HZ / 8;
205 
206 /*
207  * Time limit for merging (see comments in bfq_setup_cooperator). Set
208  * to the slowest value that, in our tests, proved to be effective in
209  * removing false positives, while not causing true positives to miss
210  * queue merging.
211  *
212  * As can be deduced from the low time limit below, queue merging, if
213  * successful, happens at the very beginning of the I/O of the involved
214  * cooperating processes, as a consequence of the arrival of the very
215  * first requests from each cooperator.  After that, there is very
216  * little chance to find cooperators.
217  */
218 static const unsigned long bfq_merge_time_limit = HZ/10;
219 
220 static struct kmem_cache *bfq_pool;
221 
222 /* Below this threshold (in ns), we consider thinktime immediate. */
223 #define BFQ_MIN_TT		(2 * NSEC_PER_MSEC)
224 
225 /* hw_tag detection: parallel requests threshold and min samples needed. */
226 #define BFQ_HW_QUEUE_THRESHOLD	3
227 #define BFQ_HW_QUEUE_SAMPLES	32
228 
229 #define BFQQ_SEEK_THR		(sector_t)(8 * 100)
230 #define BFQQ_SECT_THR_NONROT	(sector_t)(2 * 32)
231 #define BFQ_RQ_SEEKY(bfqd, last_pos, rq) \
232 	(get_sdist(last_pos, rq) >			\
233 	 BFQQ_SEEK_THR &&				\
234 	 (!blk_queue_nonrot(bfqd->queue) ||		\
235 	  blk_rq_sectors(rq) < BFQQ_SECT_THR_NONROT))
236 #define BFQQ_CLOSE_THR		(sector_t)(8 * 1024)
237 #define BFQQ_SEEKY(bfqq)	(hweight32(bfqq->seek_history) > 19)
238 /*
239  * Sync random I/O is likely to be confused with soft real-time I/O,
240  * because it is characterized by limited throughput and apparently
241  * isochronous arrival pattern. To avoid false positives, queues
242  * containing only random (seeky) I/O are prevented from being tagged
243  * as soft real-time.
244  */
245 #define BFQQ_TOTALLY_SEEKY(bfqq)	(bfqq->seek_history == -1)
246 
247 /* Min number of samples required to perform peak-rate update */
248 #define BFQ_RATE_MIN_SAMPLES	32
249 /* Min observation time interval required to perform a peak-rate update (ns) */
250 #define BFQ_RATE_MIN_INTERVAL	(300*NSEC_PER_MSEC)
251 /* Target observation time interval for a peak-rate update (ns) */
252 #define BFQ_RATE_REF_INTERVAL	NSEC_PER_SEC
253 
254 /*
255  * Shift used for peak-rate fixed precision calculations.
256  * With
257  * - the current shift: 16 positions
258  * - the current type used to store rate: u32
259  * - the current unit of measure for rate: [sectors/usec], or, more precisely,
260  *   [(sectors/usec) / 2^BFQ_RATE_SHIFT] to take into account the shift,
261  * the range of rates that can be stored is
262  * [1 / 2^BFQ_RATE_SHIFT, 2^(32 - BFQ_RATE_SHIFT)] sectors/usec =
263  * [1 / 2^16, 2^16] sectors/usec = [15e-6, 65536] sectors/usec =
264  * [15, 65G] sectors/sec
265  * Which, assuming a sector size of 512B, corresponds to a range of
266  * [7.5K, 33T] B/sec
267  */
268 #define BFQ_RATE_SHIFT		16
269 
270 /*
271  * When configured for computing the duration of the weight-raising
272  * for interactive queues automatically (see the comments at the
273  * beginning of this file), BFQ does it using the following formula:
274  * duration = (ref_rate / r) * ref_wr_duration,
275  * where r is the peak rate of the device, and ref_rate and
276  * ref_wr_duration are two reference parameters.  In particular,
277  * ref_rate is the peak rate of the reference storage device (see
278  * below), and ref_wr_duration is about the maximum time needed, with
279  * BFQ and while reading two files in parallel, to load typical large
280  * applications on the reference device (see the comments on
281  * max_service_from_wr below, for more details on how ref_wr_duration
282  * is obtained).  In practice, the slower/faster the device at hand
283  * is, the more/less it takes to load applications with respect to the
284  * reference device.  Accordingly, the longer/shorter BFQ grants
285  * weight raising to interactive applications.
286  *
287  * BFQ uses two different reference pairs (ref_rate, ref_wr_duration),
288  * depending on whether the device is rotational or non-rotational.
289  *
290  * In the following definitions, ref_rate[0] and ref_wr_duration[0]
291  * are the reference values for a rotational device, whereas
292  * ref_rate[1] and ref_wr_duration[1] are the reference values for a
293  * non-rotational device. The reference rates are not the actual peak
294  * rates of the devices used as a reference, but slightly lower
295  * values. The reason for using slightly lower values is that the
296  * peak-rate estimator tends to yield slightly lower values than the
297  * actual peak rate (it can yield the actual peak rate only if there
298  * is only one process doing I/O, and the process does sequential
299  * I/O).
300  *
301  * The reference peak rates are measured in sectors/usec, left-shifted
302  * by BFQ_RATE_SHIFT.
303  */
304 static int ref_rate[2] = {14000, 33000};
305 /*
306  * To improve readability, a conversion function is used to initialize
307  * the following array, which entails that the array can be
308  * initialized only in a function.
309  */
310 static int ref_wr_duration[2];
311 
312 /*
313  * BFQ uses the above-detailed, time-based weight-raising mechanism to
314  * privilege interactive tasks. This mechanism is vulnerable to the
315  * following false positives: I/O-bound applications that will go on
316  * doing I/O for much longer than the duration of weight
317  * raising. These applications have basically no benefit from being
318  * weight-raised at the beginning of their I/O. On the opposite end,
319  * while being weight-raised, these applications
320  * a) unjustly steal throughput to applications that may actually need
321  * low latency;
322  * b) make BFQ uselessly perform device idling; device idling results
323  * in loss of device throughput with most flash-based storage, and may
324  * increase latencies when used purposelessly.
325  *
326  * BFQ tries to reduce these problems, by adopting the following
327  * countermeasure. To introduce this countermeasure, we need first to
328  * finish explaining how the duration of weight-raising for
329  * interactive tasks is computed.
330  *
331  * For a bfq_queue deemed as interactive, the duration of weight
332  * raising is dynamically adjusted, as a function of the estimated
333  * peak rate of the device, so as to be equal to the time needed to
334  * execute the 'largest' interactive task we benchmarked so far. By
335  * largest task, we mean the task for which each involved process has
336  * to do more I/O than for any of the other tasks we benchmarked. This
337  * reference interactive task is the start-up of LibreOffice Writer,
338  * and in this task each process/bfq_queue needs to have at most ~110K
339  * sectors transferred.
340  *
341  * This last piece of information enables BFQ to reduce the actual
342  * duration of weight-raising for at least one class of I/O-bound
343  * applications: those doing sequential or quasi-sequential I/O. An
344  * example is file copy. In fact, once started, the main I/O-bound
345  * processes of these applications usually consume the above 110K
346  * sectors in much less time than the processes of an application that
347  * is starting, because these I/O-bound processes will greedily devote
348  * almost all their CPU cycles only to their target,
349  * throughput-friendly I/O operations. This is even more true if BFQ
350  * happens to be underestimating the device peak rate, and thus
351  * overestimating the duration of weight raising. But, according to
352  * our measurements, once transferred 110K sectors, these processes
353  * have no right to be weight-raised any longer.
354  *
355  * Basing on the last consideration, BFQ ends weight-raising for a
356  * bfq_queue if the latter happens to have received an amount of
357  * service at least equal to the following constant. The constant is
358  * set to slightly more than 110K, to have a minimum safety margin.
359  *
360  * This early ending of weight-raising reduces the amount of time
361  * during which interactive false positives cause the two problems
362  * described at the beginning of these comments.
363  */
364 static const unsigned long max_service_from_wr = 120000;
365 
366 /*
367  * Maximum time between the creation of two queues, for stable merge
368  * to be activated (in ms)
369  */
370 static const unsigned long bfq_activation_stable_merging = 600;
371 /*
372  * Minimum time to be waited before evaluating delayed stable merge (in ms)
373  */
374 static const unsigned long bfq_late_stable_merging = 600;
375 
376 #define RQ_BIC(rq)		((struct bfq_io_cq *)((rq)->elv.priv[0]))
377 #define RQ_BFQQ(rq)		((rq)->elv.priv[1])
378 
379 struct bfq_queue *bic_to_bfqq(struct bfq_io_cq *bic, bool is_sync,
380 			      unsigned int actuator_idx)
381 {
382 	if (is_sync)
383 		return bic->bfqq[1][actuator_idx];
384 
385 	return bic->bfqq[0][actuator_idx];
386 }
387 
388 static void bfq_put_stable_ref(struct bfq_queue *bfqq);
389 
390 void bic_set_bfqq(struct bfq_io_cq *bic,
391 		  struct bfq_queue *bfqq,
392 		  bool is_sync,
393 		  unsigned int actuator_idx)
394 {
395 	struct bfq_queue *old_bfqq = bic->bfqq[is_sync][actuator_idx];
396 
397 	/*
398 	 * If bfqq != NULL, then a non-stable queue merge between
399 	 * bic->bfqq and bfqq is happening here. This causes troubles
400 	 * in the following case: bic->bfqq has also been scheduled
401 	 * for a possible stable merge with bic->stable_merge_bfqq,
402 	 * and bic->stable_merge_bfqq == bfqq happens to
403 	 * hold. Troubles occur because bfqq may then undergo a split,
404 	 * thereby becoming eligible for a stable merge. Yet, if
405 	 * bic->stable_merge_bfqq points exactly to bfqq, then bfqq
406 	 * would be stably merged with itself. To avoid this anomaly,
407 	 * we cancel the stable merge if
408 	 * bic->stable_merge_bfqq == bfqq.
409 	 */
410 	struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[actuator_idx];
411 
412 	/* Clear bic pointer if bfqq is detached from this bic */
413 	if (old_bfqq && old_bfqq->bic == bic)
414 		old_bfqq->bic = NULL;
415 
416 	if (is_sync)
417 		bic->bfqq[1][actuator_idx] = bfqq;
418 	else
419 		bic->bfqq[0][actuator_idx] = bfqq;
420 
421 	if (bfqq && bfqq_data->stable_merge_bfqq == bfqq) {
422 		/*
423 		 * Actually, these same instructions are executed also
424 		 * in bfq_setup_cooperator, in case of abort or actual
425 		 * execution of a stable merge. We could avoid
426 		 * repeating these instructions there too, but if we
427 		 * did so, we would nest even more complexity in this
428 		 * function.
429 		 */
430 		bfq_put_stable_ref(bfqq_data->stable_merge_bfqq);
431 
432 		bfqq_data->stable_merge_bfqq = NULL;
433 	}
434 }
435 
436 struct bfq_data *bic_to_bfqd(struct bfq_io_cq *bic)
437 {
438 	return bic->icq.q->elevator->elevator_data;
439 }
440 
441 /**
442  * icq_to_bic - convert iocontext queue structure to bfq_io_cq.
443  * @icq: the iocontext queue.
444  */
445 static struct bfq_io_cq *icq_to_bic(struct io_cq *icq)
446 {
447 	/* bic->icq is the first member, %NULL will convert to %NULL */
448 	return container_of(icq, struct bfq_io_cq, icq);
449 }
450 
451 /**
452  * bfq_bic_lookup - search into @ioc a bic associated to @bfqd.
453  * @q: the request queue.
454  */
455 static struct bfq_io_cq *bfq_bic_lookup(struct request_queue *q)
456 {
457 	struct bfq_io_cq *icq;
458 	unsigned long flags;
459 
460 	if (!current->io_context)
461 		return NULL;
462 
463 	spin_lock_irqsave(&q->queue_lock, flags);
464 	icq = icq_to_bic(ioc_lookup_icq(q));
465 	spin_unlock_irqrestore(&q->queue_lock, flags);
466 
467 	return icq;
468 }
469 
470 /*
471  * Scheduler run of queue, if there are requests pending and no one in the
472  * driver that will restart queueing.
473  */
474 void bfq_schedule_dispatch(struct bfq_data *bfqd)
475 {
476 	lockdep_assert_held(&bfqd->lock);
477 
478 	if (bfqd->queued != 0) {
479 		bfq_log(bfqd, "schedule dispatch");
480 		blk_mq_run_hw_queues(bfqd->queue, true);
481 	}
482 }
483 
484 #define bfq_class_idle(bfqq)	((bfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
485 
486 #define bfq_sample_valid(samples)	((samples) > 80)
487 
488 /*
489  * Lifted from AS - choose which of rq1 and rq2 that is best served now.
490  * We choose the request that is closer to the head right now.  Distance
491  * behind the head is penalized and only allowed to a certain extent.
492  */
493 static struct request *bfq_choose_req(struct bfq_data *bfqd,
494 				      struct request *rq1,
495 				      struct request *rq2,
496 				      sector_t last)
497 {
498 	sector_t s1, s2, d1 = 0, d2 = 0;
499 	unsigned long back_max;
500 #define BFQ_RQ1_WRAP	0x01 /* request 1 wraps */
501 #define BFQ_RQ2_WRAP	0x02 /* request 2 wraps */
502 	unsigned int wrap = 0; /* bit mask: requests behind the disk head? */
503 
504 	if (!rq1 || rq1 == rq2)
505 		return rq2;
506 	if (!rq2)
507 		return rq1;
508 
509 	if (rq_is_sync(rq1) && !rq_is_sync(rq2))
510 		return rq1;
511 	else if (rq_is_sync(rq2) && !rq_is_sync(rq1))
512 		return rq2;
513 	if ((rq1->cmd_flags & REQ_META) && !(rq2->cmd_flags & REQ_META))
514 		return rq1;
515 	else if ((rq2->cmd_flags & REQ_META) && !(rq1->cmd_flags & REQ_META))
516 		return rq2;
517 
518 	s1 = blk_rq_pos(rq1);
519 	s2 = blk_rq_pos(rq2);
520 
521 	/*
522 	 * By definition, 1KiB is 2 sectors.
523 	 */
524 	back_max = bfqd->bfq_back_max * 2;
525 
526 	/*
527 	 * Strict one way elevator _except_ in the case where we allow
528 	 * short backward seeks which are biased as twice the cost of a
529 	 * similar forward seek.
530 	 */
531 	if (s1 >= last)
532 		d1 = s1 - last;
533 	else if (s1 + back_max >= last)
534 		d1 = (last - s1) * bfqd->bfq_back_penalty;
535 	else
536 		wrap |= BFQ_RQ1_WRAP;
537 
538 	if (s2 >= last)
539 		d2 = s2 - last;
540 	else if (s2 + back_max >= last)
541 		d2 = (last - s2) * bfqd->bfq_back_penalty;
542 	else
543 		wrap |= BFQ_RQ2_WRAP;
544 
545 	/* Found required data */
546 
547 	/*
548 	 * By doing switch() on the bit mask "wrap" we avoid having to
549 	 * check two variables for all permutations: --> faster!
550 	 */
551 	switch (wrap) {
552 	case 0: /* common case for CFQ: rq1 and rq2 not wrapped */
553 		if (d1 < d2)
554 			return rq1;
555 		else if (d2 < d1)
556 			return rq2;
557 
558 		if (s1 >= s2)
559 			return rq1;
560 		else
561 			return rq2;
562 
563 	case BFQ_RQ2_WRAP:
564 		return rq1;
565 	case BFQ_RQ1_WRAP:
566 		return rq2;
567 	case BFQ_RQ1_WRAP|BFQ_RQ2_WRAP: /* both rqs wrapped */
568 	default:
569 		/*
570 		 * Since both rqs are wrapped,
571 		 * start with the one that's further behind head
572 		 * (--> only *one* back seek required),
573 		 * since back seek takes more time than forward.
574 		 */
575 		if (s1 <= s2)
576 			return rq1;
577 		else
578 			return rq2;
579 	}
580 }
581 
582 #define BFQ_LIMIT_INLINE_DEPTH 16
583 
584 #ifdef CONFIG_BFQ_GROUP_IOSCHED
585 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
586 {
587 	struct bfq_data *bfqd = bfqq->bfqd;
588 	struct bfq_entity *entity = &bfqq->entity;
589 	struct bfq_entity *inline_entities[BFQ_LIMIT_INLINE_DEPTH];
590 	struct bfq_entity **entities = inline_entities;
591 	int depth, level, alloc_depth = BFQ_LIMIT_INLINE_DEPTH;
592 	int class_idx = bfqq->ioprio_class - 1;
593 	struct bfq_sched_data *sched_data;
594 	unsigned long wsum;
595 	bool ret = false;
596 
597 	if (!entity->on_st_or_in_serv)
598 		return false;
599 
600 retry:
601 	spin_lock_irq(&bfqd->lock);
602 	/* +1 for bfqq entity, root cgroup not included */
603 	depth = bfqg_to_blkg(bfqq_group(bfqq))->blkcg->css.cgroup->level + 1;
604 	if (depth > alloc_depth) {
605 		spin_unlock_irq(&bfqd->lock);
606 		if (entities != inline_entities)
607 			kfree(entities);
608 		entities = kmalloc_array(depth, sizeof(*entities), GFP_NOIO);
609 		if (!entities)
610 			return false;
611 		alloc_depth = depth;
612 		goto retry;
613 	}
614 
615 	sched_data = entity->sched_data;
616 	/* Gather our ancestors as we need to traverse them in reverse order */
617 	level = 0;
618 	for_each_entity(entity) {
619 		/*
620 		 * If at some level entity is not even active, allow request
621 		 * queueing so that BFQ knows there's work to do and activate
622 		 * entities.
623 		 */
624 		if (!entity->on_st_or_in_serv)
625 			goto out;
626 		/* Uh, more parents than cgroup subsystem thinks? */
627 		if (WARN_ON_ONCE(level >= depth))
628 			break;
629 		entities[level++] = entity;
630 	}
631 	WARN_ON_ONCE(level != depth);
632 	for (level--; level >= 0; level--) {
633 		entity = entities[level];
634 		if (level > 0) {
635 			wsum = bfq_entity_service_tree(entity)->wsum;
636 		} else {
637 			int i;
638 			/*
639 			 * For bfqq itself we take into account service trees
640 			 * of all higher priority classes and multiply their
641 			 * weights so that low prio queue from higher class
642 			 * gets more requests than high prio queue from lower
643 			 * class.
644 			 */
645 			wsum = 0;
646 			for (i = 0; i <= class_idx; i++) {
647 				wsum = wsum * IOPRIO_BE_NR +
648 					sched_data->service_tree[i].wsum;
649 			}
650 		}
651 		if (!wsum)
652 			continue;
653 		limit = DIV_ROUND_CLOSEST(limit * entity->weight, wsum);
654 		if (entity->allocated >= limit) {
655 			bfq_log_bfqq(bfqq->bfqd, bfqq,
656 				"too many requests: allocated %d limit %d level %d",
657 				entity->allocated, limit, level);
658 			ret = true;
659 			break;
660 		}
661 	}
662 out:
663 	spin_unlock_irq(&bfqd->lock);
664 	if (entities != inline_entities)
665 		kfree(entities);
666 	return ret;
667 }
668 #else
669 static bool bfqq_request_over_limit(struct bfq_queue *bfqq, int limit)
670 {
671 	return false;
672 }
673 #endif
674 
675 /*
676  * Async I/O can easily starve sync I/O (both sync reads and sync
677  * writes), by consuming all tags. Similarly, storms of sync writes,
678  * such as those that sync(2) may trigger, can starve sync reads.
679  * Limit depths of async I/O and sync writes so as to counter both
680  * problems.
681  *
682  * Also if a bfq queue or its parent cgroup consume more tags than would be
683  * appropriate for their weight, we trim the available tag depth to 1. This
684  * avoids a situation where one cgroup can starve another cgroup from tags and
685  * thus block service differentiation among cgroups. Note that because the
686  * queue / cgroup already has many requests allocated and queued, this does not
687  * significantly affect service guarantees coming from the BFQ scheduling
688  * algorithm.
689  */
690 static void bfq_limit_depth(blk_opf_t opf, struct blk_mq_alloc_data *data)
691 {
692 	struct bfq_data *bfqd = data->q->elevator->elevator_data;
693 	struct bfq_io_cq *bic = bfq_bic_lookup(data->q);
694 	int depth;
695 	unsigned limit = data->q->nr_requests;
696 	unsigned int act_idx;
697 
698 	/* Sync reads have full depth available */
699 	if (op_is_sync(opf) && !op_is_write(opf)) {
700 		depth = 0;
701 	} else {
702 		depth = bfqd->word_depths[!!bfqd->wr_busy_queues][op_is_sync(opf)];
703 		limit = (limit * depth) >> bfqd->full_depth_shift;
704 	}
705 
706 	for (act_idx = 0; bic && act_idx < bfqd->num_actuators; act_idx++) {
707 		struct bfq_queue *bfqq =
708 			bic_to_bfqq(bic, op_is_sync(opf), act_idx);
709 
710 		/*
711 		 * Does queue (or any parent entity) exceed number of
712 		 * requests that should be available to it? Heavily
713 		 * limit depth so that it cannot consume more
714 		 * available requests and thus starve other entities.
715 		 */
716 		if (bfqq && bfqq_request_over_limit(bfqq, limit)) {
717 			depth = 1;
718 			break;
719 		}
720 	}
721 	bfq_log(bfqd, "[%s] wr_busy %d sync %d depth %u",
722 		__func__, bfqd->wr_busy_queues, op_is_sync(opf), depth);
723 	if (depth)
724 		data->shallow_depth = depth;
725 }
726 
727 static struct bfq_queue *
728 bfq_rq_pos_tree_lookup(struct bfq_data *bfqd, struct rb_root *root,
729 		     sector_t sector, struct rb_node **ret_parent,
730 		     struct rb_node ***rb_link)
731 {
732 	struct rb_node **p, *parent;
733 	struct bfq_queue *bfqq = NULL;
734 
735 	parent = NULL;
736 	p = &root->rb_node;
737 	while (*p) {
738 		struct rb_node **n;
739 
740 		parent = *p;
741 		bfqq = rb_entry(parent, struct bfq_queue, pos_node);
742 
743 		/*
744 		 * Sort strictly based on sector. Smallest to the left,
745 		 * largest to the right.
746 		 */
747 		if (sector > blk_rq_pos(bfqq->next_rq))
748 			n = &(*p)->rb_right;
749 		else if (sector < blk_rq_pos(bfqq->next_rq))
750 			n = &(*p)->rb_left;
751 		else
752 			break;
753 		p = n;
754 		bfqq = NULL;
755 	}
756 
757 	*ret_parent = parent;
758 	if (rb_link)
759 		*rb_link = p;
760 
761 	bfq_log(bfqd, "rq_pos_tree_lookup %llu: returning %d",
762 		(unsigned long long)sector,
763 		bfqq ? bfqq->pid : 0);
764 
765 	return bfqq;
766 }
767 
768 static bool bfq_too_late_for_merging(struct bfq_queue *bfqq)
769 {
770 	return bfqq->service_from_backlogged > 0 &&
771 		time_is_before_jiffies(bfqq->first_IO_time +
772 				       bfq_merge_time_limit);
773 }
774 
775 /*
776  * The following function is not marked as __cold because it is
777  * actually cold, but for the same performance goal described in the
778  * comments on the likely() at the beginning of
779  * bfq_setup_cooperator(). Unexpectedly, to reach an even lower
780  * execution time for the case where this function is not invoked, we
781  * had to add an unlikely() in each involved if().
782  */
783 void __cold
784 bfq_pos_tree_add_move(struct bfq_data *bfqd, struct bfq_queue *bfqq)
785 {
786 	struct rb_node **p, *parent;
787 	struct bfq_queue *__bfqq;
788 
789 	if (bfqq->pos_root) {
790 		rb_erase(&bfqq->pos_node, bfqq->pos_root);
791 		bfqq->pos_root = NULL;
792 	}
793 
794 	/* oom_bfqq does not participate in queue merging */
795 	if (bfqq == &bfqd->oom_bfqq)
796 		return;
797 
798 	/*
799 	 * bfqq cannot be merged any longer (see comments in
800 	 * bfq_setup_cooperator): no point in adding bfqq into the
801 	 * position tree.
802 	 */
803 	if (bfq_too_late_for_merging(bfqq))
804 		return;
805 
806 	if (bfq_class_idle(bfqq))
807 		return;
808 	if (!bfqq->next_rq)
809 		return;
810 
811 	bfqq->pos_root = &bfqq_group(bfqq)->rq_pos_tree;
812 	__bfqq = bfq_rq_pos_tree_lookup(bfqd, bfqq->pos_root,
813 			blk_rq_pos(bfqq->next_rq), &parent, &p);
814 	if (!__bfqq) {
815 		rb_link_node(&bfqq->pos_node, parent, p);
816 		rb_insert_color(&bfqq->pos_node, bfqq->pos_root);
817 	} else
818 		bfqq->pos_root = NULL;
819 }
820 
821 /*
822  * The following function returns false either if every active queue
823  * must receive the same share of the throughput (symmetric scenario),
824  * or, as a special case, if bfqq must receive a share of the
825  * throughput lower than or equal to the share that every other active
826  * queue must receive.  If bfqq does sync I/O, then these are the only
827  * two cases where bfqq happens to be guaranteed its share of the
828  * throughput even if I/O dispatching is not plugged when bfqq remains
829  * temporarily empty (for more details, see the comments in the
830  * function bfq_better_to_idle()). For this reason, the return value
831  * of this function is used to check whether I/O-dispatch plugging can
832  * be avoided.
833  *
834  * The above first case (symmetric scenario) occurs when:
835  * 1) all active queues have the same weight,
836  * 2) all active queues belong to the same I/O-priority class,
837  * 3) all active groups at the same level in the groups tree have the same
838  *    weight,
839  * 4) all active groups at the same level in the groups tree have the same
840  *    number of children.
841  *
842  * Unfortunately, keeping the necessary state for evaluating exactly
843  * the last two symmetry sub-conditions above would be quite complex
844  * and time consuming. Therefore this function evaluates, instead,
845  * only the following stronger three sub-conditions, for which it is
846  * much easier to maintain the needed state:
847  * 1) all active queues have the same weight,
848  * 2) all active queues belong to the same I/O-priority class,
849  * 3) there is at most one active group.
850  * In particular, the last condition is always true if hierarchical
851  * support or the cgroups interface are not enabled, thus no state
852  * needs to be maintained in this case.
853  */
854 static bool bfq_asymmetric_scenario(struct bfq_data *bfqd,
855 				   struct bfq_queue *bfqq)
856 {
857 	bool smallest_weight = bfqq &&
858 		bfqq->weight_counter &&
859 		bfqq->weight_counter ==
860 		container_of(
861 			rb_first_cached(&bfqd->queue_weights_tree),
862 			struct bfq_weight_counter,
863 			weights_node);
864 
865 	/*
866 	 * For queue weights to differ, queue_weights_tree must contain
867 	 * at least two nodes.
868 	 */
869 	bool varied_queue_weights = !smallest_weight &&
870 		!RB_EMPTY_ROOT(&bfqd->queue_weights_tree.rb_root) &&
871 		(bfqd->queue_weights_tree.rb_root.rb_node->rb_left ||
872 		 bfqd->queue_weights_tree.rb_root.rb_node->rb_right);
873 
874 	bool multiple_classes_busy =
875 		(bfqd->busy_queues[0] && bfqd->busy_queues[1]) ||
876 		(bfqd->busy_queues[0] && bfqd->busy_queues[2]) ||
877 		(bfqd->busy_queues[1] && bfqd->busy_queues[2]);
878 
879 	return varied_queue_weights || multiple_classes_busy
880 #ifdef CONFIG_BFQ_GROUP_IOSCHED
881 	       || bfqd->num_groups_with_pending_reqs > 1
882 #endif
883 		;
884 }
885 
886 /*
887  * If the weight-counter tree passed as input contains no counter for
888  * the weight of the input queue, then add that counter; otherwise just
889  * increment the existing counter.
890  *
891  * Note that weight-counter trees contain few nodes in mostly symmetric
892  * scenarios. For example, if all queues have the same weight, then the
893  * weight-counter tree for the queues may contain at most one node.
894  * This holds even if low_latency is on, because weight-raised queues
895  * are not inserted in the tree.
896  * In most scenarios, the rate at which nodes are created/destroyed
897  * should be low too.
898  */
899 void bfq_weights_tree_add(struct bfq_queue *bfqq)
900 {
901 	struct rb_root_cached *root = &bfqq->bfqd->queue_weights_tree;
902 	struct bfq_entity *entity = &bfqq->entity;
903 	struct rb_node **new = &(root->rb_root.rb_node), *parent = NULL;
904 	bool leftmost = true;
905 
906 	/*
907 	 * Do not insert if the queue is already associated with a
908 	 * counter, which happens if:
909 	 *   1) a request arrival has caused the queue to become both
910 	 *      non-weight-raised, and hence change its weight, and
911 	 *      backlogged; in this respect, each of the two events
912 	 *      causes an invocation of this function,
913 	 *   2) this is the invocation of this function caused by the
914 	 *      second event. This second invocation is actually useless,
915 	 *      and we handle this fact by exiting immediately. More
916 	 *      efficient or clearer solutions might possibly be adopted.
917 	 */
918 	if (bfqq->weight_counter)
919 		return;
920 
921 	while (*new) {
922 		struct bfq_weight_counter *__counter = container_of(*new,
923 						struct bfq_weight_counter,
924 						weights_node);
925 		parent = *new;
926 
927 		if (entity->weight == __counter->weight) {
928 			bfqq->weight_counter = __counter;
929 			goto inc_counter;
930 		}
931 		if (entity->weight < __counter->weight)
932 			new = &((*new)->rb_left);
933 		else {
934 			new = &((*new)->rb_right);
935 			leftmost = false;
936 		}
937 	}
938 
939 	bfqq->weight_counter = kzalloc(sizeof(struct bfq_weight_counter),
940 				       GFP_ATOMIC);
941 
942 	/*
943 	 * In the unlucky event of an allocation failure, we just
944 	 * exit. This will cause the weight of queue to not be
945 	 * considered in bfq_asymmetric_scenario, which, in its turn,
946 	 * causes the scenario to be deemed wrongly symmetric in case
947 	 * bfqq's weight would have been the only weight making the
948 	 * scenario asymmetric.  On the bright side, no unbalance will
949 	 * however occur when bfqq becomes inactive again (the
950 	 * invocation of this function is triggered by an activation
951 	 * of queue).  In fact, bfq_weights_tree_remove does nothing
952 	 * if !bfqq->weight_counter.
953 	 */
954 	if (unlikely(!bfqq->weight_counter))
955 		return;
956 
957 	bfqq->weight_counter->weight = entity->weight;
958 	rb_link_node(&bfqq->weight_counter->weights_node, parent, new);
959 	rb_insert_color_cached(&bfqq->weight_counter->weights_node, root,
960 				leftmost);
961 
962 inc_counter:
963 	bfqq->weight_counter->num_active++;
964 	bfqq->ref++;
965 }
966 
967 /*
968  * Decrement the weight counter associated with the queue, and, if the
969  * counter reaches 0, remove the counter from the tree.
970  * See the comments to the function bfq_weights_tree_add() for considerations
971  * about overhead.
972  */
973 void bfq_weights_tree_remove(struct bfq_queue *bfqq)
974 {
975 	struct rb_root_cached *root;
976 
977 	if (!bfqq->weight_counter)
978 		return;
979 
980 	root = &bfqq->bfqd->queue_weights_tree;
981 	bfqq->weight_counter->num_active--;
982 	if (bfqq->weight_counter->num_active > 0)
983 		goto reset_entity_pointer;
984 
985 	rb_erase_cached(&bfqq->weight_counter->weights_node, root);
986 	kfree(bfqq->weight_counter);
987 
988 reset_entity_pointer:
989 	bfqq->weight_counter = NULL;
990 	bfq_put_queue(bfqq);
991 }
992 
993 /*
994  * Return expired entry, or NULL to just start from scratch in rbtree.
995  */
996 static struct request *bfq_check_fifo(struct bfq_queue *bfqq,
997 				      struct request *last)
998 {
999 	struct request *rq;
1000 
1001 	if (bfq_bfqq_fifo_expire(bfqq))
1002 		return NULL;
1003 
1004 	bfq_mark_bfqq_fifo_expire(bfqq);
1005 
1006 	rq = rq_entry_fifo(bfqq->fifo.next);
1007 
1008 	if (rq == last || ktime_get_ns() < rq->fifo_time)
1009 		return NULL;
1010 
1011 	bfq_log_bfqq(bfqq->bfqd, bfqq, "check_fifo: returned %p", rq);
1012 	return rq;
1013 }
1014 
1015 static struct request *bfq_find_next_rq(struct bfq_data *bfqd,
1016 					struct bfq_queue *bfqq,
1017 					struct request *last)
1018 {
1019 	struct rb_node *rbnext = rb_next(&last->rb_node);
1020 	struct rb_node *rbprev = rb_prev(&last->rb_node);
1021 	struct request *next, *prev = NULL;
1022 
1023 	/* Follow expired path, else get first next available. */
1024 	next = bfq_check_fifo(bfqq, last);
1025 	if (next)
1026 		return next;
1027 
1028 	if (rbprev)
1029 		prev = rb_entry_rq(rbprev);
1030 
1031 	if (rbnext)
1032 		next = rb_entry_rq(rbnext);
1033 	else {
1034 		rbnext = rb_first(&bfqq->sort_list);
1035 		if (rbnext && rbnext != &last->rb_node)
1036 			next = rb_entry_rq(rbnext);
1037 	}
1038 
1039 	return bfq_choose_req(bfqd, next, prev, blk_rq_pos(last));
1040 }
1041 
1042 /* see the definition of bfq_async_charge_factor for details */
1043 static unsigned long bfq_serv_to_charge(struct request *rq,
1044 					struct bfq_queue *bfqq)
1045 {
1046 	if (bfq_bfqq_sync(bfqq) || bfqq->wr_coeff > 1 ||
1047 	    bfq_asymmetric_scenario(bfqq->bfqd, bfqq))
1048 		return blk_rq_sectors(rq);
1049 
1050 	return blk_rq_sectors(rq) * bfq_async_charge_factor;
1051 }
1052 
1053 /**
1054  * bfq_updated_next_req - update the queue after a new next_rq selection.
1055  * @bfqd: the device data the queue belongs to.
1056  * @bfqq: the queue to update.
1057  *
1058  * If the first request of a queue changes we make sure that the queue
1059  * has enough budget to serve at least its first request (if the
1060  * request has grown).  We do this because if the queue has not enough
1061  * budget for its first request, it has to go through two dispatch
1062  * rounds to actually get it dispatched.
1063  */
1064 static void bfq_updated_next_req(struct bfq_data *bfqd,
1065 				 struct bfq_queue *bfqq)
1066 {
1067 	struct bfq_entity *entity = &bfqq->entity;
1068 	struct request *next_rq = bfqq->next_rq;
1069 	unsigned long new_budget;
1070 
1071 	if (!next_rq)
1072 		return;
1073 
1074 	if (bfqq == bfqd->in_service_queue)
1075 		/*
1076 		 * In order not to break guarantees, budgets cannot be
1077 		 * changed after an entity has been selected.
1078 		 */
1079 		return;
1080 
1081 	new_budget = max_t(unsigned long,
1082 			   max_t(unsigned long, bfqq->max_budget,
1083 				 bfq_serv_to_charge(next_rq, bfqq)),
1084 			   entity->service);
1085 	if (entity->budget != new_budget) {
1086 		entity->budget = new_budget;
1087 		bfq_log_bfqq(bfqd, bfqq, "updated next rq: new budget %lu",
1088 					 new_budget);
1089 		bfq_requeue_bfqq(bfqd, bfqq, false);
1090 	}
1091 }
1092 
1093 static unsigned int bfq_wr_duration(struct bfq_data *bfqd)
1094 {
1095 	u64 dur;
1096 
1097 	dur = bfqd->rate_dur_prod;
1098 	do_div(dur, bfqd->peak_rate);
1099 
1100 	/*
1101 	 * Limit duration between 3 and 25 seconds. The upper limit
1102 	 * has been conservatively set after the following worst case:
1103 	 * on a QEMU/KVM virtual machine
1104 	 * - running in a slow PC
1105 	 * - with a virtual disk stacked on a slow low-end 5400rpm HDD
1106 	 * - serving a heavy I/O workload, such as the sequential reading
1107 	 *   of several files
1108 	 * mplayer took 23 seconds to start, if constantly weight-raised.
1109 	 *
1110 	 * As for higher values than that accommodating the above bad
1111 	 * scenario, tests show that higher values would often yield
1112 	 * the opposite of the desired result, i.e., would worsen
1113 	 * responsiveness by allowing non-interactive applications to
1114 	 * preserve weight raising for too long.
1115 	 *
1116 	 * On the other end, lower values than 3 seconds make it
1117 	 * difficult for most interactive tasks to complete their jobs
1118 	 * before weight-raising finishes.
1119 	 */
1120 	return clamp_val(dur, msecs_to_jiffies(3000), msecs_to_jiffies(25000));
1121 }
1122 
1123 /* switch back from soft real-time to interactive weight raising */
1124 static void switch_back_to_interactive_wr(struct bfq_queue *bfqq,
1125 					  struct bfq_data *bfqd)
1126 {
1127 	bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1128 	bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1129 	bfqq->last_wr_start_finish = bfqq->wr_start_at_switch_to_srt;
1130 }
1131 
1132 static void
1133 bfq_bfqq_resume_state(struct bfq_queue *bfqq, struct bfq_data *bfqd,
1134 		      struct bfq_io_cq *bic, bool bfq_already_existing)
1135 {
1136 	unsigned int old_wr_coeff = 1;
1137 	bool busy = bfq_already_existing && bfq_bfqq_busy(bfqq);
1138 	unsigned int a_idx = bfqq->actuator_idx;
1139 	struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
1140 
1141 	if (bfqq_data->saved_has_short_ttime)
1142 		bfq_mark_bfqq_has_short_ttime(bfqq);
1143 	else
1144 		bfq_clear_bfqq_has_short_ttime(bfqq);
1145 
1146 	if (bfqq_data->saved_IO_bound)
1147 		bfq_mark_bfqq_IO_bound(bfqq);
1148 	else
1149 		bfq_clear_bfqq_IO_bound(bfqq);
1150 
1151 	bfqq->last_serv_time_ns = bfqq_data->saved_last_serv_time_ns;
1152 	bfqq->inject_limit = bfqq_data->saved_inject_limit;
1153 	bfqq->decrease_time_jif = bfqq_data->saved_decrease_time_jif;
1154 
1155 	bfqq->entity.new_weight = bfqq_data->saved_weight;
1156 	bfqq->ttime = bfqq_data->saved_ttime;
1157 	bfqq->io_start_time = bfqq_data->saved_io_start_time;
1158 	bfqq->tot_idle_time = bfqq_data->saved_tot_idle_time;
1159 	/*
1160 	 * Restore weight coefficient only if low_latency is on
1161 	 */
1162 	if (bfqd->low_latency) {
1163 		old_wr_coeff = bfqq->wr_coeff;
1164 		bfqq->wr_coeff = bfqq_data->saved_wr_coeff;
1165 	}
1166 	bfqq->service_from_wr = bfqq_data->saved_service_from_wr;
1167 	bfqq->wr_start_at_switch_to_srt =
1168 		bfqq_data->saved_wr_start_at_switch_to_srt;
1169 	bfqq->last_wr_start_finish = bfqq_data->saved_last_wr_start_finish;
1170 	bfqq->wr_cur_max_time = bfqq_data->saved_wr_cur_max_time;
1171 
1172 	if (bfqq->wr_coeff > 1 && (bfq_bfqq_in_large_burst(bfqq) ||
1173 	    time_is_before_jiffies(bfqq->last_wr_start_finish +
1174 				   bfqq->wr_cur_max_time))) {
1175 		if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
1176 		    !bfq_bfqq_in_large_burst(bfqq) &&
1177 		    time_is_after_eq_jiffies(bfqq->wr_start_at_switch_to_srt +
1178 					     bfq_wr_duration(bfqd))) {
1179 			switch_back_to_interactive_wr(bfqq, bfqd);
1180 		} else {
1181 			bfqq->wr_coeff = 1;
1182 			bfq_log_bfqq(bfqq->bfqd, bfqq,
1183 				     "resume state: switching off wr");
1184 		}
1185 	}
1186 
1187 	/* make sure weight will be updated, however we got here */
1188 	bfqq->entity.prio_changed = 1;
1189 
1190 	if (likely(!busy))
1191 		return;
1192 
1193 	if (old_wr_coeff == 1 && bfqq->wr_coeff > 1)
1194 		bfqd->wr_busy_queues++;
1195 	else if (old_wr_coeff > 1 && bfqq->wr_coeff == 1)
1196 		bfqd->wr_busy_queues--;
1197 }
1198 
1199 static int bfqq_process_refs(struct bfq_queue *bfqq)
1200 {
1201 	return bfqq->ref - bfqq->entity.allocated -
1202 		bfqq->entity.on_st_or_in_serv -
1203 		(bfqq->weight_counter != NULL) - bfqq->stable_ref;
1204 }
1205 
1206 /* Empty burst list and add just bfqq (see comments on bfq_handle_burst) */
1207 static void bfq_reset_burst_list(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1208 {
1209 	struct bfq_queue *item;
1210 	struct hlist_node *n;
1211 
1212 	hlist_for_each_entry_safe(item, n, &bfqd->burst_list, burst_list_node)
1213 		hlist_del_init(&item->burst_list_node);
1214 
1215 	/*
1216 	 * Start the creation of a new burst list only if there is no
1217 	 * active queue. See comments on the conditional invocation of
1218 	 * bfq_handle_burst().
1219 	 */
1220 	if (bfq_tot_busy_queues(bfqd) == 0) {
1221 		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1222 		bfqd->burst_size = 1;
1223 	} else
1224 		bfqd->burst_size = 0;
1225 
1226 	bfqd->burst_parent_entity = bfqq->entity.parent;
1227 }
1228 
1229 /* Add bfqq to the list of queues in current burst (see bfq_handle_burst) */
1230 static void bfq_add_to_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1231 {
1232 	/* Increment burst size to take into account also bfqq */
1233 	bfqd->burst_size++;
1234 
1235 	if (bfqd->burst_size == bfqd->bfq_large_burst_thresh) {
1236 		struct bfq_queue *pos, *bfqq_item;
1237 		struct hlist_node *n;
1238 
1239 		/*
1240 		 * Enough queues have been activated shortly after each
1241 		 * other to consider this burst as large.
1242 		 */
1243 		bfqd->large_burst = true;
1244 
1245 		/*
1246 		 * We can now mark all queues in the burst list as
1247 		 * belonging to a large burst.
1248 		 */
1249 		hlist_for_each_entry(bfqq_item, &bfqd->burst_list,
1250 				     burst_list_node)
1251 			bfq_mark_bfqq_in_large_burst(bfqq_item);
1252 		bfq_mark_bfqq_in_large_burst(bfqq);
1253 
1254 		/*
1255 		 * From now on, and until the current burst finishes, any
1256 		 * new queue being activated shortly after the last queue
1257 		 * was inserted in the burst can be immediately marked as
1258 		 * belonging to a large burst. So the burst list is not
1259 		 * needed any more. Remove it.
1260 		 */
1261 		hlist_for_each_entry_safe(pos, n, &bfqd->burst_list,
1262 					  burst_list_node)
1263 			hlist_del_init(&pos->burst_list_node);
1264 	} else /*
1265 		* Burst not yet large: add bfqq to the burst list. Do
1266 		* not increment the ref counter for bfqq, because bfqq
1267 		* is removed from the burst list before freeing bfqq
1268 		* in put_queue.
1269 		*/
1270 		hlist_add_head(&bfqq->burst_list_node, &bfqd->burst_list);
1271 }
1272 
1273 /*
1274  * If many queues belonging to the same group happen to be created
1275  * shortly after each other, then the processes associated with these
1276  * queues have typically a common goal. In particular, bursts of queue
1277  * creations are usually caused by services or applications that spawn
1278  * many parallel threads/processes. Examples are systemd during boot,
1279  * or git grep. To help these processes get their job done as soon as
1280  * possible, it is usually better to not grant either weight-raising
1281  * or device idling to their queues, unless these queues must be
1282  * protected from the I/O flowing through other active queues.
1283  *
1284  * In this comment we describe, firstly, the reasons why this fact
1285  * holds, and, secondly, the next function, which implements the main
1286  * steps needed to properly mark these queues so that they can then be
1287  * treated in a different way.
1288  *
1289  * The above services or applications benefit mostly from a high
1290  * throughput: the quicker the requests of the activated queues are
1291  * cumulatively served, the sooner the target job of these queues gets
1292  * completed. As a consequence, weight-raising any of these queues,
1293  * which also implies idling the device for it, is almost always
1294  * counterproductive, unless there are other active queues to isolate
1295  * these new queues from. If there no other active queues, then
1296  * weight-raising these new queues just lowers throughput in most
1297  * cases.
1298  *
1299  * On the other hand, a burst of queue creations may be caused also by
1300  * the start of an application that does not consist of a lot of
1301  * parallel I/O-bound threads. In fact, with a complex application,
1302  * several short processes may need to be executed to start-up the
1303  * application. In this respect, to start an application as quickly as
1304  * possible, the best thing to do is in any case to privilege the I/O
1305  * related to the application with respect to all other
1306  * I/O. Therefore, the best strategy to start as quickly as possible
1307  * an application that causes a burst of queue creations is to
1308  * weight-raise all the queues created during the burst. This is the
1309  * exact opposite of the best strategy for the other type of bursts.
1310  *
1311  * In the end, to take the best action for each of the two cases, the
1312  * two types of bursts need to be distinguished. Fortunately, this
1313  * seems relatively easy, by looking at the sizes of the bursts. In
1314  * particular, we found a threshold such that only bursts with a
1315  * larger size than that threshold are apparently caused by
1316  * services or commands such as systemd or git grep. For brevity,
1317  * hereafter we call just 'large' these bursts. BFQ *does not*
1318  * weight-raise queues whose creation occurs in a large burst. In
1319  * addition, for each of these queues BFQ performs or does not perform
1320  * idling depending on which choice boosts the throughput more. The
1321  * exact choice depends on the device and request pattern at
1322  * hand.
1323  *
1324  * Unfortunately, false positives may occur while an interactive task
1325  * is starting (e.g., an application is being started). The
1326  * consequence is that the queues associated with the task do not
1327  * enjoy weight raising as expected. Fortunately these false positives
1328  * are very rare. They typically occur if some service happens to
1329  * start doing I/O exactly when the interactive task starts.
1330  *
1331  * Turning back to the next function, it is invoked only if there are
1332  * no active queues (apart from active queues that would belong to the
1333  * same, possible burst bfqq would belong to), and it implements all
1334  * the steps needed to detect the occurrence of a large burst and to
1335  * properly mark all the queues belonging to it (so that they can then
1336  * be treated in a different way). This goal is achieved by
1337  * maintaining a "burst list" that holds, temporarily, the queues that
1338  * belong to the burst in progress. The list is then used to mark
1339  * these queues as belonging to a large burst if the burst does become
1340  * large. The main steps are the following.
1341  *
1342  * . when the very first queue is created, the queue is inserted into the
1343  *   list (as it could be the first queue in a possible burst)
1344  *
1345  * . if the current burst has not yet become large, and a queue Q that does
1346  *   not yet belong to the burst is activated shortly after the last time
1347  *   at which a new queue entered the burst list, then the function appends
1348  *   Q to the burst list
1349  *
1350  * . if, as a consequence of the previous step, the burst size reaches
1351  *   the large-burst threshold, then
1352  *
1353  *     . all the queues in the burst list are marked as belonging to a
1354  *       large burst
1355  *
1356  *     . the burst list is deleted; in fact, the burst list already served
1357  *       its purpose (keeping temporarily track of the queues in a burst,
1358  *       so as to be able to mark them as belonging to a large burst in the
1359  *       previous sub-step), and now is not needed any more
1360  *
1361  *     . the device enters a large-burst mode
1362  *
1363  * . if a queue Q that does not belong to the burst is created while
1364  *   the device is in large-burst mode and shortly after the last time
1365  *   at which a queue either entered the burst list or was marked as
1366  *   belonging to the current large burst, then Q is immediately marked
1367  *   as belonging to a large burst.
1368  *
1369  * . if a queue Q that does not belong to the burst is created a while
1370  *   later, i.e., not shortly after, than the last time at which a queue
1371  *   either entered the burst list or was marked as belonging to the
1372  *   current large burst, then the current burst is deemed as finished and:
1373  *
1374  *        . the large-burst mode is reset if set
1375  *
1376  *        . the burst list is emptied
1377  *
1378  *        . Q is inserted in the burst list, as Q may be the first queue
1379  *          in a possible new burst (then the burst list contains just Q
1380  *          after this step).
1381  */
1382 static void bfq_handle_burst(struct bfq_data *bfqd, struct bfq_queue *bfqq)
1383 {
1384 	/*
1385 	 * If bfqq is already in the burst list or is part of a large
1386 	 * burst, or finally has just been split, then there is
1387 	 * nothing else to do.
1388 	 */
1389 	if (!hlist_unhashed(&bfqq->burst_list_node) ||
1390 	    bfq_bfqq_in_large_burst(bfqq) ||
1391 	    time_is_after_eq_jiffies(bfqq->split_time +
1392 				     msecs_to_jiffies(10)))
1393 		return;
1394 
1395 	/*
1396 	 * If bfqq's creation happens late enough, or bfqq belongs to
1397 	 * a different group than the burst group, then the current
1398 	 * burst is finished, and related data structures must be
1399 	 * reset.
1400 	 *
1401 	 * In this respect, consider the special case where bfqq is
1402 	 * the very first queue created after BFQ is selected for this
1403 	 * device. In this case, last_ins_in_burst and
1404 	 * burst_parent_entity are not yet significant when we get
1405 	 * here. But it is easy to verify that, whether or not the
1406 	 * following condition is true, bfqq will end up being
1407 	 * inserted into the burst list. In particular the list will
1408 	 * happen to contain only bfqq. And this is exactly what has
1409 	 * to happen, as bfqq may be the first queue of the first
1410 	 * burst.
1411 	 */
1412 	if (time_is_before_jiffies(bfqd->last_ins_in_burst +
1413 	    bfqd->bfq_burst_interval) ||
1414 	    bfqq->entity.parent != bfqd->burst_parent_entity) {
1415 		bfqd->large_burst = false;
1416 		bfq_reset_burst_list(bfqd, bfqq);
1417 		goto end;
1418 	}
1419 
1420 	/*
1421 	 * If we get here, then bfqq is being activated shortly after the
1422 	 * last queue. So, if the current burst is also large, we can mark
1423 	 * bfqq as belonging to this large burst immediately.
1424 	 */
1425 	if (bfqd->large_burst) {
1426 		bfq_mark_bfqq_in_large_burst(bfqq);
1427 		goto end;
1428 	}
1429 
1430 	/*
1431 	 * If we get here, then a large-burst state has not yet been
1432 	 * reached, but bfqq is being activated shortly after the last
1433 	 * queue. Then we add bfqq to the burst.
1434 	 */
1435 	bfq_add_to_burst(bfqd, bfqq);
1436 end:
1437 	/*
1438 	 * At this point, bfqq either has been added to the current
1439 	 * burst or has caused the current burst to terminate and a
1440 	 * possible new burst to start. In particular, in the second
1441 	 * case, bfqq has become the first queue in the possible new
1442 	 * burst.  In both cases last_ins_in_burst needs to be moved
1443 	 * forward.
1444 	 */
1445 	bfqd->last_ins_in_burst = jiffies;
1446 }
1447 
1448 static int bfq_bfqq_budget_left(struct bfq_queue *bfqq)
1449 {
1450 	struct bfq_entity *entity = &bfqq->entity;
1451 
1452 	return entity->budget - entity->service;
1453 }
1454 
1455 /*
1456  * If enough samples have been computed, return the current max budget
1457  * stored in bfqd, which is dynamically updated according to the
1458  * estimated disk peak rate; otherwise return the default max budget
1459  */
1460 static int bfq_max_budget(struct bfq_data *bfqd)
1461 {
1462 	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1463 		return bfq_default_max_budget;
1464 	else
1465 		return bfqd->bfq_max_budget;
1466 }
1467 
1468 /*
1469  * Return min budget, which is a fraction of the current or default
1470  * max budget (trying with 1/32)
1471  */
1472 static int bfq_min_budget(struct bfq_data *bfqd)
1473 {
1474 	if (bfqd->budgets_assigned < bfq_stats_min_budgets)
1475 		return bfq_default_max_budget / 32;
1476 	else
1477 		return bfqd->bfq_max_budget / 32;
1478 }
1479 
1480 /*
1481  * The next function, invoked after the input queue bfqq switches from
1482  * idle to busy, updates the budget of bfqq. The function also tells
1483  * whether the in-service queue should be expired, by returning
1484  * true. The purpose of expiring the in-service queue is to give bfqq
1485  * the chance to possibly preempt the in-service queue, and the reason
1486  * for preempting the in-service queue is to achieve one of the two
1487  * goals below.
1488  *
1489  * 1. Guarantee to bfqq its reserved bandwidth even if bfqq has
1490  * expired because it has remained idle. In particular, bfqq may have
1491  * expired for one of the following two reasons:
1492  *
1493  * - BFQQE_NO_MORE_REQUESTS bfqq did not enjoy any device idling
1494  *   and did not make it to issue a new request before its last
1495  *   request was served;
1496  *
1497  * - BFQQE_TOO_IDLE bfqq did enjoy device idling, but did not issue
1498  *   a new request before the expiration of the idling-time.
1499  *
1500  * Even if bfqq has expired for one of the above reasons, the process
1501  * associated with the queue may be however issuing requests greedily,
1502  * and thus be sensitive to the bandwidth it receives (bfqq may have
1503  * remained idle for other reasons: CPU high load, bfqq not enjoying
1504  * idling, I/O throttling somewhere in the path from the process to
1505  * the I/O scheduler, ...). But if, after every expiration for one of
1506  * the above two reasons, bfqq has to wait for the service of at least
1507  * one full budget of another queue before being served again, then
1508  * bfqq is likely to get a much lower bandwidth or resource time than
1509  * its reserved ones. To address this issue, two countermeasures need
1510  * to be taken.
1511  *
1512  * First, the budget and the timestamps of bfqq need to be updated in
1513  * a special way on bfqq reactivation: they need to be updated as if
1514  * bfqq did not remain idle and did not expire. In fact, if they are
1515  * computed as if bfqq expired and remained idle until reactivation,
1516  * then the process associated with bfqq is treated as if, instead of
1517  * being greedy, it stopped issuing requests when bfqq remained idle,
1518  * and restarts issuing requests only on this reactivation. In other
1519  * words, the scheduler does not help the process recover the "service
1520  * hole" between bfqq expiration and reactivation. As a consequence,
1521  * the process receives a lower bandwidth than its reserved one. In
1522  * contrast, to recover this hole, the budget must be updated as if
1523  * bfqq was not expired at all before this reactivation, i.e., it must
1524  * be set to the value of the remaining budget when bfqq was
1525  * expired. Along the same line, timestamps need to be assigned the
1526  * value they had the last time bfqq was selected for service, i.e.,
1527  * before last expiration. Thus timestamps need to be back-shifted
1528  * with respect to their normal computation (see [1] for more details
1529  * on this tricky aspect).
1530  *
1531  * Secondly, to allow the process to recover the hole, the in-service
1532  * queue must be expired too, to give bfqq the chance to preempt it
1533  * immediately. In fact, if bfqq has to wait for a full budget of the
1534  * in-service queue to be completed, then it may become impossible to
1535  * let the process recover the hole, even if the back-shifted
1536  * timestamps of bfqq are lower than those of the in-service queue. If
1537  * this happens for most or all of the holes, then the process may not
1538  * receive its reserved bandwidth. In this respect, it is worth noting
1539  * that, being the service of outstanding requests unpreemptible, a
1540  * little fraction of the holes may however be unrecoverable, thereby
1541  * causing a little loss of bandwidth.
1542  *
1543  * The last important point is detecting whether bfqq does need this
1544  * bandwidth recovery. In this respect, the next function deems the
1545  * process associated with bfqq greedy, and thus allows it to recover
1546  * the hole, if: 1) the process is waiting for the arrival of a new
1547  * request (which implies that bfqq expired for one of the above two
1548  * reasons), and 2) such a request has arrived soon. The first
1549  * condition is controlled through the flag non_blocking_wait_rq,
1550  * while the second through the flag arrived_in_time. If both
1551  * conditions hold, then the function computes the budget in the
1552  * above-described special way, and signals that the in-service queue
1553  * should be expired. Timestamp back-shifting is done later in
1554  * __bfq_activate_entity.
1555  *
1556  * 2. Reduce latency. Even if timestamps are not backshifted to let
1557  * the process associated with bfqq recover a service hole, bfqq may
1558  * however happen to have, after being (re)activated, a lower finish
1559  * timestamp than the in-service queue.	 That is, the next budget of
1560  * bfqq may have to be completed before the one of the in-service
1561  * queue. If this is the case, then preempting the in-service queue
1562  * allows this goal to be achieved, apart from the unpreemptible,
1563  * outstanding requests mentioned above.
1564  *
1565  * Unfortunately, regardless of which of the above two goals one wants
1566  * to achieve, service trees need first to be updated to know whether
1567  * the in-service queue must be preempted. To have service trees
1568  * correctly updated, the in-service queue must be expired and
1569  * rescheduled, and bfqq must be scheduled too. This is one of the
1570  * most costly operations (in future versions, the scheduling
1571  * mechanism may be re-designed in such a way to make it possible to
1572  * know whether preemption is needed without needing to update service
1573  * trees). In addition, queue preemptions almost always cause random
1574  * I/O, which may in turn cause loss of throughput. Finally, there may
1575  * even be no in-service queue when the next function is invoked (so,
1576  * no queue to compare timestamps with). Because of these facts, the
1577  * next function adopts the following simple scheme to avoid costly
1578  * operations, too frequent preemptions and too many dependencies on
1579  * the state of the scheduler: it requests the expiration of the
1580  * in-service queue (unconditionally) only for queues that need to
1581  * recover a hole. Then it delegates to other parts of the code the
1582  * responsibility of handling the above case 2.
1583  */
1584 static bool bfq_bfqq_update_budg_for_activation(struct bfq_data *bfqd,
1585 						struct bfq_queue *bfqq,
1586 						bool arrived_in_time)
1587 {
1588 	struct bfq_entity *entity = &bfqq->entity;
1589 
1590 	/*
1591 	 * In the next compound condition, we check also whether there
1592 	 * is some budget left, because otherwise there is no point in
1593 	 * trying to go on serving bfqq with this same budget: bfqq
1594 	 * would be expired immediately after being selected for
1595 	 * service. This would only cause useless overhead.
1596 	 */
1597 	if (bfq_bfqq_non_blocking_wait_rq(bfqq) && arrived_in_time &&
1598 	    bfq_bfqq_budget_left(bfqq) > 0) {
1599 		/*
1600 		 * We do not clear the flag non_blocking_wait_rq here, as
1601 		 * the latter is used in bfq_activate_bfqq to signal
1602 		 * that timestamps need to be back-shifted (and is
1603 		 * cleared right after).
1604 		 */
1605 
1606 		/*
1607 		 * In next assignment we rely on that either
1608 		 * entity->service or entity->budget are not updated
1609 		 * on expiration if bfqq is empty (see
1610 		 * __bfq_bfqq_recalc_budget). Thus both quantities
1611 		 * remain unchanged after such an expiration, and the
1612 		 * following statement therefore assigns to
1613 		 * entity->budget the remaining budget on such an
1614 		 * expiration.
1615 		 */
1616 		entity->budget = min_t(unsigned long,
1617 				       bfq_bfqq_budget_left(bfqq),
1618 				       bfqq->max_budget);
1619 
1620 		/*
1621 		 * At this point, we have used entity->service to get
1622 		 * the budget left (needed for updating
1623 		 * entity->budget). Thus we finally can, and have to,
1624 		 * reset entity->service. The latter must be reset
1625 		 * because bfqq would otherwise be charged again for
1626 		 * the service it has received during its previous
1627 		 * service slot(s).
1628 		 */
1629 		entity->service = 0;
1630 
1631 		return true;
1632 	}
1633 
1634 	/*
1635 	 * We can finally complete expiration, by setting service to 0.
1636 	 */
1637 	entity->service = 0;
1638 	entity->budget = max_t(unsigned long, bfqq->max_budget,
1639 			       bfq_serv_to_charge(bfqq->next_rq, bfqq));
1640 	bfq_clear_bfqq_non_blocking_wait_rq(bfqq);
1641 	return false;
1642 }
1643 
1644 /*
1645  * Return the farthest past time instant according to jiffies
1646  * macros.
1647  */
1648 static unsigned long bfq_smallest_from_now(void)
1649 {
1650 	return jiffies - MAX_JIFFY_OFFSET;
1651 }
1652 
1653 static void bfq_update_bfqq_wr_on_rq_arrival(struct bfq_data *bfqd,
1654 					     struct bfq_queue *bfqq,
1655 					     unsigned int old_wr_coeff,
1656 					     bool wr_or_deserves_wr,
1657 					     bool interactive,
1658 					     bool in_burst,
1659 					     bool soft_rt)
1660 {
1661 	if (old_wr_coeff == 1 && wr_or_deserves_wr) {
1662 		/* start a weight-raising period */
1663 		if (interactive) {
1664 			bfqq->service_from_wr = 0;
1665 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1666 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1667 		} else {
1668 			/*
1669 			 * No interactive weight raising in progress
1670 			 * here: assign minus infinity to
1671 			 * wr_start_at_switch_to_srt, to make sure
1672 			 * that, at the end of the soft-real-time
1673 			 * weight raising periods that is starting
1674 			 * now, no interactive weight-raising period
1675 			 * may be wrongly considered as still in
1676 			 * progress (and thus actually started by
1677 			 * mistake).
1678 			 */
1679 			bfqq->wr_start_at_switch_to_srt =
1680 				bfq_smallest_from_now();
1681 			bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1682 				BFQ_SOFTRT_WEIGHT_FACTOR;
1683 			bfqq->wr_cur_max_time =
1684 				bfqd->bfq_wr_rt_max_time;
1685 		}
1686 
1687 		/*
1688 		 * If needed, further reduce budget to make sure it is
1689 		 * close to bfqq's backlog, so as to reduce the
1690 		 * scheduling-error component due to a too large
1691 		 * budget. Do not care about throughput consequences,
1692 		 * but only about latency. Finally, do not assign a
1693 		 * too small budget either, to avoid increasing
1694 		 * latency by causing too frequent expirations.
1695 		 */
1696 		bfqq->entity.budget = min_t(unsigned long,
1697 					    bfqq->entity.budget,
1698 					    2 * bfq_min_budget(bfqd));
1699 	} else if (old_wr_coeff > 1) {
1700 		if (interactive) { /* update wr coeff and duration */
1701 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
1702 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
1703 		} else if (in_burst)
1704 			bfqq->wr_coeff = 1;
1705 		else if (soft_rt) {
1706 			/*
1707 			 * The application is now or still meeting the
1708 			 * requirements for being deemed soft rt.  We
1709 			 * can then correctly and safely (re)charge
1710 			 * the weight-raising duration for the
1711 			 * application with the weight-raising
1712 			 * duration for soft rt applications.
1713 			 *
1714 			 * In particular, doing this recharge now, i.e.,
1715 			 * before the weight-raising period for the
1716 			 * application finishes, reduces the probability
1717 			 * of the following negative scenario:
1718 			 * 1) the weight of a soft rt application is
1719 			 *    raised at startup (as for any newly
1720 			 *    created application),
1721 			 * 2) since the application is not interactive,
1722 			 *    at a certain time weight-raising is
1723 			 *    stopped for the application,
1724 			 * 3) at that time the application happens to
1725 			 *    still have pending requests, and hence
1726 			 *    is destined to not have a chance to be
1727 			 *    deemed soft rt before these requests are
1728 			 *    completed (see the comments to the
1729 			 *    function bfq_bfqq_softrt_next_start()
1730 			 *    for details on soft rt detection),
1731 			 * 4) these pending requests experience a high
1732 			 *    latency because the application is not
1733 			 *    weight-raised while they are pending.
1734 			 */
1735 			if (bfqq->wr_cur_max_time !=
1736 				bfqd->bfq_wr_rt_max_time) {
1737 				bfqq->wr_start_at_switch_to_srt =
1738 					bfqq->last_wr_start_finish;
1739 
1740 				bfqq->wr_cur_max_time =
1741 					bfqd->bfq_wr_rt_max_time;
1742 				bfqq->wr_coeff = bfqd->bfq_wr_coeff *
1743 					BFQ_SOFTRT_WEIGHT_FACTOR;
1744 			}
1745 			bfqq->last_wr_start_finish = jiffies;
1746 		}
1747 	}
1748 }
1749 
1750 static bool bfq_bfqq_idle_for_long_time(struct bfq_data *bfqd,
1751 					struct bfq_queue *bfqq)
1752 {
1753 	return bfqq->dispatched == 0 &&
1754 		time_is_before_jiffies(
1755 			bfqq->budget_timeout +
1756 			bfqd->bfq_wr_min_idle_time);
1757 }
1758 
1759 
1760 /*
1761  * Return true if bfqq is in a higher priority class, or has a higher
1762  * weight than the in-service queue.
1763  */
1764 static bool bfq_bfqq_higher_class_or_weight(struct bfq_queue *bfqq,
1765 					    struct bfq_queue *in_serv_bfqq)
1766 {
1767 	int bfqq_weight, in_serv_weight;
1768 
1769 	if (bfqq->ioprio_class < in_serv_bfqq->ioprio_class)
1770 		return true;
1771 
1772 	if (in_serv_bfqq->entity.parent == bfqq->entity.parent) {
1773 		bfqq_weight = bfqq->entity.weight;
1774 		in_serv_weight = in_serv_bfqq->entity.weight;
1775 	} else {
1776 		if (bfqq->entity.parent)
1777 			bfqq_weight = bfqq->entity.parent->weight;
1778 		else
1779 			bfqq_weight = bfqq->entity.weight;
1780 		if (in_serv_bfqq->entity.parent)
1781 			in_serv_weight = in_serv_bfqq->entity.parent->weight;
1782 		else
1783 			in_serv_weight = in_serv_bfqq->entity.weight;
1784 	}
1785 
1786 	return bfqq_weight > in_serv_weight;
1787 }
1788 
1789 /*
1790  * Get the index of the actuator that will serve bio.
1791  */
1792 static unsigned int bfq_actuator_index(struct bfq_data *bfqd, struct bio *bio)
1793 {
1794 	unsigned int i;
1795 	sector_t end;
1796 
1797 	/* no search needed if one or zero ranges present */
1798 	if (bfqd->num_actuators == 1)
1799 		return 0;
1800 
1801 	/* bio_end_sector(bio) gives the sector after the last one */
1802 	end = bio_end_sector(bio) - 1;
1803 
1804 	for (i = 0; i < bfqd->num_actuators; i++) {
1805 		if (end >= bfqd->sector[i] &&
1806 		    end < bfqd->sector[i] + bfqd->nr_sectors[i])
1807 			return i;
1808 	}
1809 
1810 	WARN_ONCE(true,
1811 		  "bfq_actuator_index: bio sector out of ranges: end=%llu\n",
1812 		  end);
1813 	return 0;
1814 }
1815 
1816 static bool bfq_better_to_idle(struct bfq_queue *bfqq);
1817 
1818 static void bfq_bfqq_handle_idle_busy_switch(struct bfq_data *bfqd,
1819 					     struct bfq_queue *bfqq,
1820 					     int old_wr_coeff,
1821 					     struct request *rq,
1822 					     bool *interactive)
1823 {
1824 	bool soft_rt, in_burst,	wr_or_deserves_wr,
1825 		bfqq_wants_to_preempt,
1826 		idle_for_long_time = bfq_bfqq_idle_for_long_time(bfqd, bfqq),
1827 		/*
1828 		 * See the comments on
1829 		 * bfq_bfqq_update_budg_for_activation for
1830 		 * details on the usage of the next variable.
1831 		 */
1832 		arrived_in_time =  ktime_get_ns() <=
1833 			bfqq->ttime.last_end_request +
1834 			bfqd->bfq_slice_idle * 3;
1835 	unsigned int act_idx = bfq_actuator_index(bfqd, rq->bio);
1836 	bool bfqq_non_merged_or_stably_merged =
1837 		bfqq->bic || RQ_BIC(rq)->bfqq_data[act_idx].stably_merged;
1838 
1839 	/*
1840 	 * bfqq deserves to be weight-raised if:
1841 	 * - it is sync,
1842 	 * - it does not belong to a large burst,
1843 	 * - it has been idle for enough time or is soft real-time,
1844 	 * - is linked to a bfq_io_cq (it is not shared in any sense),
1845 	 * - has a default weight (otherwise we assume the user wanted
1846 	 *   to control its weight explicitly)
1847 	 */
1848 	in_burst = bfq_bfqq_in_large_burst(bfqq);
1849 	soft_rt = bfqd->bfq_wr_max_softrt_rate > 0 &&
1850 		!BFQQ_TOTALLY_SEEKY(bfqq) &&
1851 		!in_burst &&
1852 		time_is_before_jiffies(bfqq->soft_rt_next_start) &&
1853 		bfqq->dispatched == 0 &&
1854 		bfqq->entity.new_weight == 40;
1855 	*interactive = !in_burst && idle_for_long_time &&
1856 		bfqq->entity.new_weight == 40;
1857 	/*
1858 	 * Merged bfq_queues are kept out of weight-raising
1859 	 * (low-latency) mechanisms. The reason is that these queues
1860 	 * are usually created for non-interactive and
1861 	 * non-soft-real-time tasks. Yet this is not the case for
1862 	 * stably-merged queues. These queues are merged just because
1863 	 * they are created shortly after each other. So they may
1864 	 * easily serve the I/O of an interactive or soft-real time
1865 	 * application, if the application happens to spawn multiple
1866 	 * processes. So let also stably-merged queued enjoy weight
1867 	 * raising.
1868 	 */
1869 	wr_or_deserves_wr = bfqd->low_latency &&
1870 		(bfqq->wr_coeff > 1 ||
1871 		 (bfq_bfqq_sync(bfqq) && bfqq_non_merged_or_stably_merged &&
1872 		  (*interactive || soft_rt)));
1873 
1874 	/*
1875 	 * Using the last flag, update budget and check whether bfqq
1876 	 * may want to preempt the in-service queue.
1877 	 */
1878 	bfqq_wants_to_preempt =
1879 		bfq_bfqq_update_budg_for_activation(bfqd, bfqq,
1880 						    arrived_in_time);
1881 
1882 	/*
1883 	 * If bfqq happened to be activated in a burst, but has been
1884 	 * idle for much more than an interactive queue, then we
1885 	 * assume that, in the overall I/O initiated in the burst, the
1886 	 * I/O associated with bfqq is finished. So bfqq does not need
1887 	 * to be treated as a queue belonging to a burst
1888 	 * anymore. Accordingly, we reset bfqq's in_large_burst flag
1889 	 * if set, and remove bfqq from the burst list if it's
1890 	 * there. We do not decrement burst_size, because the fact
1891 	 * that bfqq does not need to belong to the burst list any
1892 	 * more does not invalidate the fact that bfqq was created in
1893 	 * a burst.
1894 	 */
1895 	if (likely(!bfq_bfqq_just_created(bfqq)) &&
1896 	    idle_for_long_time &&
1897 	    time_is_before_jiffies(
1898 		    bfqq->budget_timeout +
1899 		    msecs_to_jiffies(10000))) {
1900 		hlist_del_init(&bfqq->burst_list_node);
1901 		bfq_clear_bfqq_in_large_burst(bfqq);
1902 	}
1903 
1904 	bfq_clear_bfqq_just_created(bfqq);
1905 
1906 	if (bfqd->low_latency) {
1907 		if (unlikely(time_is_after_jiffies(bfqq->split_time)))
1908 			/* wraparound */
1909 			bfqq->split_time =
1910 				jiffies - bfqd->bfq_wr_min_idle_time - 1;
1911 
1912 		if (time_is_before_jiffies(bfqq->split_time +
1913 					   bfqd->bfq_wr_min_idle_time)) {
1914 			bfq_update_bfqq_wr_on_rq_arrival(bfqd, bfqq,
1915 							 old_wr_coeff,
1916 							 wr_or_deserves_wr,
1917 							 *interactive,
1918 							 in_burst,
1919 							 soft_rt);
1920 
1921 			if (old_wr_coeff != bfqq->wr_coeff)
1922 				bfqq->entity.prio_changed = 1;
1923 		}
1924 	}
1925 
1926 	bfqq->last_idle_bklogged = jiffies;
1927 	bfqq->service_from_backlogged = 0;
1928 	bfq_clear_bfqq_softrt_update(bfqq);
1929 
1930 	bfq_add_bfqq_busy(bfqq);
1931 
1932 	/*
1933 	 * Expire in-service queue if preemption may be needed for
1934 	 * guarantees or throughput. As for guarantees, we care
1935 	 * explicitly about two cases. The first is that bfqq has to
1936 	 * recover a service hole, as explained in the comments on
1937 	 * bfq_bfqq_update_budg_for_activation(), i.e., that
1938 	 * bfqq_wants_to_preempt is true. However, if bfqq does not
1939 	 * carry time-critical I/O, then bfqq's bandwidth is less
1940 	 * important than that of queues that carry time-critical I/O.
1941 	 * So, as a further constraint, we consider this case only if
1942 	 * bfqq is at least as weight-raised, i.e., at least as time
1943 	 * critical, as the in-service queue.
1944 	 *
1945 	 * The second case is that bfqq is in a higher priority class,
1946 	 * or has a higher weight than the in-service queue. If this
1947 	 * condition does not hold, we don't care because, even if
1948 	 * bfqq does not start to be served immediately, the resulting
1949 	 * delay for bfqq's I/O is however lower or much lower than
1950 	 * the ideal completion time to be guaranteed to bfqq's I/O.
1951 	 *
1952 	 * In both cases, preemption is needed only if, according to
1953 	 * the timestamps of both bfqq and of the in-service queue,
1954 	 * bfqq actually is the next queue to serve. So, to reduce
1955 	 * useless preemptions, the return value of
1956 	 * next_queue_may_preempt() is considered in the next compound
1957 	 * condition too. Yet next_queue_may_preempt() just checks a
1958 	 * simple, necessary condition for bfqq to be the next queue
1959 	 * to serve. In fact, to evaluate a sufficient condition, the
1960 	 * timestamps of the in-service queue would need to be
1961 	 * updated, and this operation is quite costly (see the
1962 	 * comments on bfq_bfqq_update_budg_for_activation()).
1963 	 *
1964 	 * As for throughput, we ask bfq_better_to_idle() whether we
1965 	 * still need to plug I/O dispatching. If bfq_better_to_idle()
1966 	 * says no, then plugging is not needed any longer, either to
1967 	 * boost throughput or to perserve service guarantees. Then
1968 	 * the best option is to stop plugging I/O, as not doing so
1969 	 * would certainly lower throughput. We may end up in this
1970 	 * case if: (1) upon a dispatch attempt, we detected that it
1971 	 * was better to plug I/O dispatch, and to wait for a new
1972 	 * request to arrive for the currently in-service queue, but
1973 	 * (2) this switch of bfqq to busy changes the scenario.
1974 	 */
1975 	if (bfqd->in_service_queue &&
1976 	    ((bfqq_wants_to_preempt &&
1977 	      bfqq->wr_coeff >= bfqd->in_service_queue->wr_coeff) ||
1978 	     bfq_bfqq_higher_class_or_weight(bfqq, bfqd->in_service_queue) ||
1979 	     !bfq_better_to_idle(bfqd->in_service_queue)) &&
1980 	    next_queue_may_preempt(bfqd))
1981 		bfq_bfqq_expire(bfqd, bfqd->in_service_queue,
1982 				false, BFQQE_PREEMPTED);
1983 }
1984 
1985 static void bfq_reset_inject_limit(struct bfq_data *bfqd,
1986 				   struct bfq_queue *bfqq)
1987 {
1988 	/* invalidate baseline total service time */
1989 	bfqq->last_serv_time_ns = 0;
1990 
1991 	/*
1992 	 * Reset pointer in case we are waiting for
1993 	 * some request completion.
1994 	 */
1995 	bfqd->waited_rq = NULL;
1996 
1997 	/*
1998 	 * If bfqq has a short think time, then start by setting the
1999 	 * inject limit to 0 prudentially, because the service time of
2000 	 * an injected I/O request may be higher than the think time
2001 	 * of bfqq, and therefore, if one request was injected when
2002 	 * bfqq remains empty, this injected request might delay the
2003 	 * service of the next I/O request for bfqq significantly. In
2004 	 * case bfqq can actually tolerate some injection, then the
2005 	 * adaptive update will however raise the limit soon. This
2006 	 * lucky circumstance holds exactly because bfqq has a short
2007 	 * think time, and thus, after remaining empty, is likely to
2008 	 * get new I/O enqueued---and then completed---before being
2009 	 * expired. This is the very pattern that gives the
2010 	 * limit-update algorithm the chance to measure the effect of
2011 	 * injection on request service times, and then to update the
2012 	 * limit accordingly.
2013 	 *
2014 	 * However, in the following special case, the inject limit is
2015 	 * left to 1 even if the think time is short: bfqq's I/O is
2016 	 * synchronized with that of some other queue, i.e., bfqq may
2017 	 * receive new I/O only after the I/O of the other queue is
2018 	 * completed. Keeping the inject limit to 1 allows the
2019 	 * blocking I/O to be served while bfqq is in service. And
2020 	 * this is very convenient both for bfqq and for overall
2021 	 * throughput, as explained in detail in the comments in
2022 	 * bfq_update_has_short_ttime().
2023 	 *
2024 	 * On the opposite end, if bfqq has a long think time, then
2025 	 * start directly by 1, because:
2026 	 * a) on the bright side, keeping at most one request in
2027 	 * service in the drive is unlikely to cause any harm to the
2028 	 * latency of bfqq's requests, as the service time of a single
2029 	 * request is likely to be lower than the think time of bfqq;
2030 	 * b) on the downside, after becoming empty, bfqq is likely to
2031 	 * expire before getting its next request. With this request
2032 	 * arrival pattern, it is very hard to sample total service
2033 	 * times and update the inject limit accordingly (see comments
2034 	 * on bfq_update_inject_limit()). So the limit is likely to be
2035 	 * never, or at least seldom, updated.  As a consequence, by
2036 	 * setting the limit to 1, we avoid that no injection ever
2037 	 * occurs with bfqq. On the downside, this proactive step
2038 	 * further reduces chances to actually compute the baseline
2039 	 * total service time. Thus it reduces chances to execute the
2040 	 * limit-update algorithm and possibly raise the limit to more
2041 	 * than 1.
2042 	 */
2043 	if (bfq_bfqq_has_short_ttime(bfqq))
2044 		bfqq->inject_limit = 0;
2045 	else
2046 		bfqq->inject_limit = 1;
2047 
2048 	bfqq->decrease_time_jif = jiffies;
2049 }
2050 
2051 static void bfq_update_io_intensity(struct bfq_queue *bfqq, u64 now_ns)
2052 {
2053 	u64 tot_io_time = now_ns - bfqq->io_start_time;
2054 
2055 	if (RB_EMPTY_ROOT(&bfqq->sort_list) && bfqq->dispatched == 0)
2056 		bfqq->tot_idle_time +=
2057 			now_ns - bfqq->ttime.last_end_request;
2058 
2059 	if (unlikely(bfq_bfqq_just_created(bfqq)))
2060 		return;
2061 
2062 	/*
2063 	 * Must be busy for at least about 80% of the time to be
2064 	 * considered I/O bound.
2065 	 */
2066 	if (bfqq->tot_idle_time * 5 > tot_io_time)
2067 		bfq_clear_bfqq_IO_bound(bfqq);
2068 	else
2069 		bfq_mark_bfqq_IO_bound(bfqq);
2070 
2071 	/*
2072 	 * Keep an observation window of at most 200 ms in the past
2073 	 * from now.
2074 	 */
2075 	if (tot_io_time > 200 * NSEC_PER_MSEC) {
2076 		bfqq->io_start_time = now_ns - (tot_io_time>>1);
2077 		bfqq->tot_idle_time >>= 1;
2078 	}
2079 }
2080 
2081 /*
2082  * Detect whether bfqq's I/O seems synchronized with that of some
2083  * other queue, i.e., whether bfqq, after remaining empty, happens to
2084  * receive new I/O only right after some I/O request of the other
2085  * queue has been completed. We call waker queue the other queue, and
2086  * we assume, for simplicity, that bfqq may have at most one waker
2087  * queue.
2088  *
2089  * A remarkable throughput boost can be reached by unconditionally
2090  * injecting the I/O of the waker queue, every time a new
2091  * bfq_dispatch_request happens to be invoked while I/O is being
2092  * plugged for bfqq.  In addition to boosting throughput, this
2093  * unblocks bfqq's I/O, thereby improving bandwidth and latency for
2094  * bfqq. Note that these same results may be achieved with the general
2095  * injection mechanism, but less effectively. For details on this
2096  * aspect, see the comments on the choice of the queue for injection
2097  * in bfq_select_queue().
2098  *
2099  * Turning back to the detection of a waker queue, a queue Q is deemed as a
2100  * waker queue for bfqq if, for three consecutive times, bfqq happens to become
2101  * non empty right after a request of Q has been completed within given
2102  * timeout. In this respect, even if bfqq is empty, we do not check for a waker
2103  * if it still has some in-flight I/O. In fact, in this case bfqq is actually
2104  * still being served by the drive, and may receive new I/O on the completion
2105  * of some of the in-flight requests. In particular, on the first time, Q is
2106  * tentatively set as a candidate waker queue, while on the third consecutive
2107  * time that Q is detected, the field waker_bfqq is set to Q, to confirm that Q
2108  * is a waker queue for bfqq. These detection steps are performed only if bfqq
2109  * has a long think time, so as to make it more likely that bfqq's I/O is
2110  * actually being blocked by a synchronization. This last filter, plus the
2111  * above three-times requirement and time limit for detection, make false
2112  * positives less likely.
2113  *
2114  * NOTE
2115  *
2116  * The sooner a waker queue is detected, the sooner throughput can be
2117  * boosted by injecting I/O from the waker queue. Fortunately,
2118  * detection is likely to be actually fast, for the following
2119  * reasons. While blocked by synchronization, bfqq has a long think
2120  * time. This implies that bfqq's inject limit is at least equal to 1
2121  * (see the comments in bfq_update_inject_limit()). So, thanks to
2122  * injection, the waker queue is likely to be served during the very
2123  * first I/O-plugging time interval for bfqq. This triggers the first
2124  * step of the detection mechanism. Thanks again to injection, the
2125  * candidate waker queue is then likely to be confirmed no later than
2126  * during the next I/O-plugging interval for bfqq.
2127  *
2128  * ISSUE
2129  *
2130  * On queue merging all waker information is lost.
2131  */
2132 static void bfq_check_waker(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2133 			    u64 now_ns)
2134 {
2135 	char waker_name[MAX_BFQQ_NAME_LENGTH];
2136 
2137 	if (!bfqd->last_completed_rq_bfqq ||
2138 	    bfqd->last_completed_rq_bfqq == bfqq ||
2139 	    bfq_bfqq_has_short_ttime(bfqq) ||
2140 	    now_ns - bfqd->last_completion >= 4 * NSEC_PER_MSEC ||
2141 	    bfqd->last_completed_rq_bfqq == &bfqd->oom_bfqq ||
2142 	    bfqq == &bfqd->oom_bfqq)
2143 		return;
2144 
2145 	/*
2146 	 * We reset waker detection logic also if too much time has passed
2147  	 * since the first detection. If wakeups are rare, pointless idling
2148 	 * doesn't hurt throughput that much. The condition below makes sure
2149 	 * we do not uselessly idle blocking waker in more than 1/64 cases.
2150 	 */
2151 	if (bfqd->last_completed_rq_bfqq !=
2152 	    bfqq->tentative_waker_bfqq ||
2153 	    now_ns > bfqq->waker_detection_started +
2154 					128 * (u64)bfqd->bfq_slice_idle) {
2155 		/*
2156 		 * First synchronization detected with a
2157 		 * candidate waker queue, or with a different
2158 		 * candidate waker queue from the current one.
2159 		 */
2160 		bfqq->tentative_waker_bfqq =
2161 			bfqd->last_completed_rq_bfqq;
2162 		bfqq->num_waker_detections = 1;
2163 		bfqq->waker_detection_started = now_ns;
2164 		bfq_bfqq_name(bfqq->tentative_waker_bfqq, waker_name,
2165 			      MAX_BFQQ_NAME_LENGTH);
2166 		bfq_log_bfqq(bfqd, bfqq, "set tentative waker %s", waker_name);
2167 	} else /* Same tentative waker queue detected again */
2168 		bfqq->num_waker_detections++;
2169 
2170 	if (bfqq->num_waker_detections == 3) {
2171 		bfqq->waker_bfqq = bfqd->last_completed_rq_bfqq;
2172 		bfqq->tentative_waker_bfqq = NULL;
2173 		bfq_bfqq_name(bfqq->waker_bfqq, waker_name,
2174 			      MAX_BFQQ_NAME_LENGTH);
2175 		bfq_log_bfqq(bfqd, bfqq, "set waker %s", waker_name);
2176 
2177 		/*
2178 		 * If the waker queue disappears, then
2179 		 * bfqq->waker_bfqq must be reset. To
2180 		 * this goal, we maintain in each
2181 		 * waker queue a list, woken_list, of
2182 		 * all the queues that reference the
2183 		 * waker queue through their
2184 		 * waker_bfqq pointer. When the waker
2185 		 * queue exits, the waker_bfqq pointer
2186 		 * of all the queues in the woken_list
2187 		 * is reset.
2188 		 *
2189 		 * In addition, if bfqq is already in
2190 		 * the woken_list of a waker queue,
2191 		 * then, before being inserted into
2192 		 * the woken_list of a new waker
2193 		 * queue, bfqq must be removed from
2194 		 * the woken_list of the old waker
2195 		 * queue.
2196 		 */
2197 		if (!hlist_unhashed(&bfqq->woken_list_node))
2198 			hlist_del_init(&bfqq->woken_list_node);
2199 		hlist_add_head(&bfqq->woken_list_node,
2200 			       &bfqd->last_completed_rq_bfqq->woken_list);
2201 	}
2202 }
2203 
2204 static void bfq_add_request(struct request *rq)
2205 {
2206 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
2207 	struct bfq_data *bfqd = bfqq->bfqd;
2208 	struct request *next_rq, *prev;
2209 	unsigned int old_wr_coeff = bfqq->wr_coeff;
2210 	bool interactive = false;
2211 	u64 now_ns = ktime_get_ns();
2212 
2213 	bfq_log_bfqq(bfqd, bfqq, "add_request %d", rq_is_sync(rq));
2214 	bfqq->queued[rq_is_sync(rq)]++;
2215 	/*
2216 	 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2217 	 * may be read without holding the lock in bfq_has_work().
2218 	 */
2219 	WRITE_ONCE(bfqd->queued, bfqd->queued + 1);
2220 
2221 	if (bfq_bfqq_sync(bfqq) && RQ_BIC(rq)->requests <= 1) {
2222 		bfq_check_waker(bfqd, bfqq, now_ns);
2223 
2224 		/*
2225 		 * Periodically reset inject limit, to make sure that
2226 		 * the latter eventually drops in case workload
2227 		 * changes, see step (3) in the comments on
2228 		 * bfq_update_inject_limit().
2229 		 */
2230 		if (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2231 					     msecs_to_jiffies(1000)))
2232 			bfq_reset_inject_limit(bfqd, bfqq);
2233 
2234 		/*
2235 		 * The following conditions must hold to setup a new
2236 		 * sampling of total service time, and then a new
2237 		 * update of the inject limit:
2238 		 * - bfqq is in service, because the total service
2239 		 *   time is evaluated only for the I/O requests of
2240 		 *   the queues in service;
2241 		 * - this is the right occasion to compute or to
2242 		 *   lower the baseline total service time, because
2243 		 *   there are actually no requests in the drive,
2244 		 *   or
2245 		 *   the baseline total service time is available, and
2246 		 *   this is the right occasion to compute the other
2247 		 *   quantity needed to update the inject limit, i.e.,
2248 		 *   the total service time caused by the amount of
2249 		 *   injection allowed by the current value of the
2250 		 *   limit. It is the right occasion because injection
2251 		 *   has actually been performed during the service
2252 		 *   hole, and there are still in-flight requests,
2253 		 *   which are very likely to be exactly the injected
2254 		 *   requests, or part of them;
2255 		 * - the minimum interval for sampling the total
2256 		 *   service time and updating the inject limit has
2257 		 *   elapsed.
2258 		 */
2259 		if (bfqq == bfqd->in_service_queue &&
2260 		    (bfqd->tot_rq_in_driver == 0 ||
2261 		     (bfqq->last_serv_time_ns > 0 &&
2262 		      bfqd->rqs_injected && bfqd->tot_rq_in_driver > 0)) &&
2263 		    time_is_before_eq_jiffies(bfqq->decrease_time_jif +
2264 					      msecs_to_jiffies(10))) {
2265 			bfqd->last_empty_occupied_ns = ktime_get_ns();
2266 			/*
2267 			 * Start the state machine for measuring the
2268 			 * total service time of rq: setting
2269 			 * wait_dispatch will cause bfqd->waited_rq to
2270 			 * be set when rq will be dispatched.
2271 			 */
2272 			bfqd->wait_dispatch = true;
2273 			/*
2274 			 * If there is no I/O in service in the drive,
2275 			 * then possible injection occurred before the
2276 			 * arrival of rq will not affect the total
2277 			 * service time of rq. So the injection limit
2278 			 * must not be updated as a function of such
2279 			 * total service time, unless new injection
2280 			 * occurs before rq is completed. To have the
2281 			 * injection limit updated only in the latter
2282 			 * case, reset rqs_injected here (rqs_injected
2283 			 * will be set in case injection is performed
2284 			 * on bfqq before rq is completed).
2285 			 */
2286 			if (bfqd->tot_rq_in_driver == 0)
2287 				bfqd->rqs_injected = false;
2288 		}
2289 	}
2290 
2291 	if (bfq_bfqq_sync(bfqq))
2292 		bfq_update_io_intensity(bfqq, now_ns);
2293 
2294 	elv_rb_add(&bfqq->sort_list, rq);
2295 
2296 	/*
2297 	 * Check if this request is a better next-serve candidate.
2298 	 */
2299 	prev = bfqq->next_rq;
2300 	next_rq = bfq_choose_req(bfqd, bfqq->next_rq, rq, bfqd->last_position);
2301 	bfqq->next_rq = next_rq;
2302 
2303 	/*
2304 	 * Adjust priority tree position, if next_rq changes.
2305 	 * See comments on bfq_pos_tree_add_move() for the unlikely().
2306 	 */
2307 	if (unlikely(!bfqd->nonrot_with_queueing && prev != bfqq->next_rq))
2308 		bfq_pos_tree_add_move(bfqd, bfqq);
2309 
2310 	if (!bfq_bfqq_busy(bfqq)) /* switching to busy ... */
2311 		bfq_bfqq_handle_idle_busy_switch(bfqd, bfqq, old_wr_coeff,
2312 						 rq, &interactive);
2313 	else {
2314 		if (bfqd->low_latency && old_wr_coeff == 1 && !rq_is_sync(rq) &&
2315 		    time_is_before_jiffies(
2316 				bfqq->last_wr_start_finish +
2317 				bfqd->bfq_wr_min_inter_arr_async)) {
2318 			bfqq->wr_coeff = bfqd->bfq_wr_coeff;
2319 			bfqq->wr_cur_max_time = bfq_wr_duration(bfqd);
2320 
2321 			bfqd->wr_busy_queues++;
2322 			bfqq->entity.prio_changed = 1;
2323 		}
2324 		if (prev != bfqq->next_rq)
2325 			bfq_updated_next_req(bfqd, bfqq);
2326 	}
2327 
2328 	/*
2329 	 * Assign jiffies to last_wr_start_finish in the following
2330 	 * cases:
2331 	 *
2332 	 * . if bfqq is not going to be weight-raised, because, for
2333 	 *   non weight-raised queues, last_wr_start_finish stores the
2334 	 *   arrival time of the last request; as of now, this piece
2335 	 *   of information is used only for deciding whether to
2336 	 *   weight-raise async queues
2337 	 *
2338 	 * . if bfqq is not weight-raised, because, if bfqq is now
2339 	 *   switching to weight-raised, then last_wr_start_finish
2340 	 *   stores the time when weight-raising starts
2341 	 *
2342 	 * . if bfqq is interactive, because, regardless of whether
2343 	 *   bfqq is currently weight-raised, the weight-raising
2344 	 *   period must start or restart (this case is considered
2345 	 *   separately because it is not detected by the above
2346 	 *   conditions, if bfqq is already weight-raised)
2347 	 *
2348 	 * last_wr_start_finish has to be updated also if bfqq is soft
2349 	 * real-time, because the weight-raising period is constantly
2350 	 * restarted on idle-to-busy transitions for these queues, but
2351 	 * this is already done in bfq_bfqq_handle_idle_busy_switch if
2352 	 * needed.
2353 	 */
2354 	if (bfqd->low_latency &&
2355 		(old_wr_coeff == 1 || bfqq->wr_coeff == 1 || interactive))
2356 		bfqq->last_wr_start_finish = jiffies;
2357 }
2358 
2359 static struct request *bfq_find_rq_fmerge(struct bfq_data *bfqd,
2360 					  struct bio *bio,
2361 					  struct request_queue *q)
2362 {
2363 	struct bfq_queue *bfqq = bfqd->bio_bfqq;
2364 
2365 
2366 	if (bfqq)
2367 		return elv_rb_find(&bfqq->sort_list, bio_end_sector(bio));
2368 
2369 	return NULL;
2370 }
2371 
2372 static sector_t get_sdist(sector_t last_pos, struct request *rq)
2373 {
2374 	if (last_pos)
2375 		return abs(blk_rq_pos(rq) - last_pos);
2376 
2377 	return 0;
2378 }
2379 
2380 static void bfq_remove_request(struct request_queue *q,
2381 			       struct request *rq)
2382 {
2383 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
2384 	struct bfq_data *bfqd = bfqq->bfqd;
2385 	const int sync = rq_is_sync(rq);
2386 
2387 	if (bfqq->next_rq == rq) {
2388 		bfqq->next_rq = bfq_find_next_rq(bfqd, bfqq, rq);
2389 		bfq_updated_next_req(bfqd, bfqq);
2390 	}
2391 
2392 	if (rq->queuelist.prev != &rq->queuelist)
2393 		list_del_init(&rq->queuelist);
2394 	bfqq->queued[sync]--;
2395 	/*
2396 	 * Updating of 'bfqd->queued' is protected by 'bfqd->lock', however, it
2397 	 * may be read without holding the lock in bfq_has_work().
2398 	 */
2399 	WRITE_ONCE(bfqd->queued, bfqd->queued - 1);
2400 	elv_rb_del(&bfqq->sort_list, rq);
2401 
2402 	elv_rqhash_del(q, rq);
2403 	if (q->last_merge == rq)
2404 		q->last_merge = NULL;
2405 
2406 	if (RB_EMPTY_ROOT(&bfqq->sort_list)) {
2407 		bfqq->next_rq = NULL;
2408 
2409 		if (bfq_bfqq_busy(bfqq) && bfqq != bfqd->in_service_queue) {
2410 			bfq_del_bfqq_busy(bfqq, false);
2411 			/*
2412 			 * bfqq emptied. In normal operation, when
2413 			 * bfqq is empty, bfqq->entity.service and
2414 			 * bfqq->entity.budget must contain,
2415 			 * respectively, the service received and the
2416 			 * budget used last time bfqq emptied. These
2417 			 * facts do not hold in this case, as at least
2418 			 * this last removal occurred while bfqq is
2419 			 * not in service. To avoid inconsistencies,
2420 			 * reset both bfqq->entity.service and
2421 			 * bfqq->entity.budget, if bfqq has still a
2422 			 * process that may issue I/O requests to it.
2423 			 */
2424 			bfqq->entity.budget = bfqq->entity.service = 0;
2425 		}
2426 
2427 		/*
2428 		 * Remove queue from request-position tree as it is empty.
2429 		 */
2430 		if (bfqq->pos_root) {
2431 			rb_erase(&bfqq->pos_node, bfqq->pos_root);
2432 			bfqq->pos_root = NULL;
2433 		}
2434 	} else {
2435 		/* see comments on bfq_pos_tree_add_move() for the unlikely() */
2436 		if (unlikely(!bfqd->nonrot_with_queueing))
2437 			bfq_pos_tree_add_move(bfqd, bfqq);
2438 	}
2439 
2440 	if (rq->cmd_flags & REQ_META)
2441 		bfqq->meta_pending--;
2442 
2443 }
2444 
2445 static bool bfq_bio_merge(struct request_queue *q, struct bio *bio,
2446 		unsigned int nr_segs)
2447 {
2448 	struct bfq_data *bfqd = q->elevator->elevator_data;
2449 	struct request *free = NULL;
2450 	/*
2451 	 * bfq_bic_lookup grabs the queue_lock: invoke it now and
2452 	 * store its return value for later use, to avoid nesting
2453 	 * queue_lock inside the bfqd->lock. We assume that the bic
2454 	 * returned by bfq_bic_lookup does not go away before
2455 	 * bfqd->lock is taken.
2456 	 */
2457 	struct bfq_io_cq *bic = bfq_bic_lookup(q);
2458 	bool ret;
2459 
2460 	spin_lock_irq(&bfqd->lock);
2461 
2462 	if (bic) {
2463 		/*
2464 		 * Make sure cgroup info is uptodate for current process before
2465 		 * considering the merge.
2466 		 */
2467 		bfq_bic_update_cgroup(bic, bio);
2468 
2469 		bfqd->bio_bfqq = bic_to_bfqq(bic, op_is_sync(bio->bi_opf),
2470 					     bfq_actuator_index(bfqd, bio));
2471 	} else {
2472 		bfqd->bio_bfqq = NULL;
2473 	}
2474 	bfqd->bio_bic = bic;
2475 
2476 	ret = blk_mq_sched_try_merge(q, bio, nr_segs, &free);
2477 
2478 	spin_unlock_irq(&bfqd->lock);
2479 	if (free)
2480 		blk_mq_free_request(free);
2481 
2482 	return ret;
2483 }
2484 
2485 static int bfq_request_merge(struct request_queue *q, struct request **req,
2486 			     struct bio *bio)
2487 {
2488 	struct bfq_data *bfqd = q->elevator->elevator_data;
2489 	struct request *__rq;
2490 
2491 	__rq = bfq_find_rq_fmerge(bfqd, bio, q);
2492 	if (__rq && elv_bio_merge_ok(__rq, bio)) {
2493 		*req = __rq;
2494 
2495 		if (blk_discard_mergable(__rq))
2496 			return ELEVATOR_DISCARD_MERGE;
2497 		return ELEVATOR_FRONT_MERGE;
2498 	}
2499 
2500 	return ELEVATOR_NO_MERGE;
2501 }
2502 
2503 static void bfq_request_merged(struct request_queue *q, struct request *req,
2504 			       enum elv_merge type)
2505 {
2506 	if (type == ELEVATOR_FRONT_MERGE &&
2507 	    rb_prev(&req->rb_node) &&
2508 	    blk_rq_pos(req) <
2509 	    blk_rq_pos(container_of(rb_prev(&req->rb_node),
2510 				    struct request, rb_node))) {
2511 		struct bfq_queue *bfqq = RQ_BFQQ(req);
2512 		struct bfq_data *bfqd;
2513 		struct request *prev, *next_rq;
2514 
2515 		if (!bfqq)
2516 			return;
2517 
2518 		bfqd = bfqq->bfqd;
2519 
2520 		/* Reposition request in its sort_list */
2521 		elv_rb_del(&bfqq->sort_list, req);
2522 		elv_rb_add(&bfqq->sort_list, req);
2523 
2524 		/* Choose next request to be served for bfqq */
2525 		prev = bfqq->next_rq;
2526 		next_rq = bfq_choose_req(bfqd, bfqq->next_rq, req,
2527 					 bfqd->last_position);
2528 		bfqq->next_rq = next_rq;
2529 		/*
2530 		 * If next_rq changes, update both the queue's budget to
2531 		 * fit the new request and the queue's position in its
2532 		 * rq_pos_tree.
2533 		 */
2534 		if (prev != bfqq->next_rq) {
2535 			bfq_updated_next_req(bfqd, bfqq);
2536 			/*
2537 			 * See comments on bfq_pos_tree_add_move() for
2538 			 * the unlikely().
2539 			 */
2540 			if (unlikely(!bfqd->nonrot_with_queueing))
2541 				bfq_pos_tree_add_move(bfqd, bfqq);
2542 		}
2543 	}
2544 }
2545 
2546 /*
2547  * This function is called to notify the scheduler that the requests
2548  * rq and 'next' have been merged, with 'next' going away.  BFQ
2549  * exploits this hook to address the following issue: if 'next' has a
2550  * fifo_time lower that rq, then the fifo_time of rq must be set to
2551  * the value of 'next', to not forget the greater age of 'next'.
2552  *
2553  * NOTE: in this function we assume that rq is in a bfq_queue, basing
2554  * on that rq is picked from the hash table q->elevator->hash, which,
2555  * in its turn, is filled only with I/O requests present in
2556  * bfq_queues, while BFQ is in use for the request queue q. In fact,
2557  * the function that fills this hash table (elv_rqhash_add) is called
2558  * only by bfq_insert_request.
2559  */
2560 static void bfq_requests_merged(struct request_queue *q, struct request *rq,
2561 				struct request *next)
2562 {
2563 	struct bfq_queue *bfqq = RQ_BFQQ(rq),
2564 		*next_bfqq = RQ_BFQQ(next);
2565 
2566 	if (!bfqq)
2567 		goto remove;
2568 
2569 	/*
2570 	 * If next and rq belong to the same bfq_queue and next is older
2571 	 * than rq, then reposition rq in the fifo (by substituting next
2572 	 * with rq). Otherwise, if next and rq belong to different
2573 	 * bfq_queues, never reposition rq: in fact, we would have to
2574 	 * reposition it with respect to next's position in its own fifo,
2575 	 * which would most certainly be too expensive with respect to
2576 	 * the benefits.
2577 	 */
2578 	if (bfqq == next_bfqq &&
2579 	    !list_empty(&rq->queuelist) && !list_empty(&next->queuelist) &&
2580 	    next->fifo_time < rq->fifo_time) {
2581 		list_del_init(&rq->queuelist);
2582 		list_replace_init(&next->queuelist, &rq->queuelist);
2583 		rq->fifo_time = next->fifo_time;
2584 	}
2585 
2586 	if (bfqq->next_rq == next)
2587 		bfqq->next_rq = rq;
2588 
2589 	bfqg_stats_update_io_merged(bfqq_group(bfqq), next->cmd_flags);
2590 remove:
2591 	/* Merged request may be in the IO scheduler. Remove it. */
2592 	if (!RB_EMPTY_NODE(&next->rb_node)) {
2593 		bfq_remove_request(next->q, next);
2594 		if (next_bfqq)
2595 			bfqg_stats_update_io_remove(bfqq_group(next_bfqq),
2596 						    next->cmd_flags);
2597 	}
2598 }
2599 
2600 /* Must be called with bfqq != NULL */
2601 static void bfq_bfqq_end_wr(struct bfq_queue *bfqq)
2602 {
2603 	/*
2604 	 * If bfqq has been enjoying interactive weight-raising, then
2605 	 * reset soft_rt_next_start. We do it for the following
2606 	 * reason. bfqq may have been conveying the I/O needed to load
2607 	 * a soft real-time application. Such an application actually
2608 	 * exhibits a soft real-time I/O pattern after it finishes
2609 	 * loading, and finally starts doing its job. But, if bfqq has
2610 	 * been receiving a lot of bandwidth so far (likely to happen
2611 	 * on a fast device), then soft_rt_next_start now contains a
2612 	 * high value that. So, without this reset, bfqq would be
2613 	 * prevented from being possibly considered as soft_rt for a
2614 	 * very long time.
2615 	 */
2616 
2617 	if (bfqq->wr_cur_max_time !=
2618 	    bfqq->bfqd->bfq_wr_rt_max_time)
2619 		bfqq->soft_rt_next_start = jiffies;
2620 
2621 	if (bfq_bfqq_busy(bfqq))
2622 		bfqq->bfqd->wr_busy_queues--;
2623 	bfqq->wr_coeff = 1;
2624 	bfqq->wr_cur_max_time = 0;
2625 	bfqq->last_wr_start_finish = jiffies;
2626 	/*
2627 	 * Trigger a weight change on the next invocation of
2628 	 * __bfq_entity_update_weight_prio.
2629 	 */
2630 	bfqq->entity.prio_changed = 1;
2631 }
2632 
2633 void bfq_end_wr_async_queues(struct bfq_data *bfqd,
2634 			     struct bfq_group *bfqg)
2635 {
2636 	int i, j, k;
2637 
2638 	for (k = 0; k < bfqd->num_actuators; k++) {
2639 		for (i = 0; i < 2; i++)
2640 			for (j = 0; j < IOPRIO_NR_LEVELS; j++)
2641 				if (bfqg->async_bfqq[i][j][k])
2642 					bfq_bfqq_end_wr(bfqg->async_bfqq[i][j][k]);
2643 		if (bfqg->async_idle_bfqq[k])
2644 			bfq_bfqq_end_wr(bfqg->async_idle_bfqq[k]);
2645 	}
2646 }
2647 
2648 static void bfq_end_wr(struct bfq_data *bfqd)
2649 {
2650 	struct bfq_queue *bfqq;
2651 	int i;
2652 
2653 	spin_lock_irq(&bfqd->lock);
2654 
2655 	for (i = 0; i < bfqd->num_actuators; i++) {
2656 		list_for_each_entry(bfqq, &bfqd->active_list[i], bfqq_list)
2657 			bfq_bfqq_end_wr(bfqq);
2658 	}
2659 	list_for_each_entry(bfqq, &bfqd->idle_list, bfqq_list)
2660 		bfq_bfqq_end_wr(bfqq);
2661 	bfq_end_wr_async(bfqd);
2662 
2663 	spin_unlock_irq(&bfqd->lock);
2664 }
2665 
2666 static sector_t bfq_io_struct_pos(void *io_struct, bool request)
2667 {
2668 	if (request)
2669 		return blk_rq_pos(io_struct);
2670 	else
2671 		return ((struct bio *)io_struct)->bi_iter.bi_sector;
2672 }
2673 
2674 static int bfq_rq_close_to_sector(void *io_struct, bool request,
2675 				  sector_t sector)
2676 {
2677 	return abs(bfq_io_struct_pos(io_struct, request) - sector) <=
2678 	       BFQQ_CLOSE_THR;
2679 }
2680 
2681 static struct bfq_queue *bfqq_find_close(struct bfq_data *bfqd,
2682 					 struct bfq_queue *bfqq,
2683 					 sector_t sector)
2684 {
2685 	struct rb_root *root = &bfqq_group(bfqq)->rq_pos_tree;
2686 	struct rb_node *parent, *node;
2687 	struct bfq_queue *__bfqq;
2688 
2689 	if (RB_EMPTY_ROOT(root))
2690 		return NULL;
2691 
2692 	/*
2693 	 * First, if we find a request starting at the end of the last
2694 	 * request, choose it.
2695 	 */
2696 	__bfqq = bfq_rq_pos_tree_lookup(bfqd, root, sector, &parent, NULL);
2697 	if (__bfqq)
2698 		return __bfqq;
2699 
2700 	/*
2701 	 * If the exact sector wasn't found, the parent of the NULL leaf
2702 	 * will contain the closest sector (rq_pos_tree sorted by
2703 	 * next_request position).
2704 	 */
2705 	__bfqq = rb_entry(parent, struct bfq_queue, pos_node);
2706 	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2707 		return __bfqq;
2708 
2709 	if (blk_rq_pos(__bfqq->next_rq) < sector)
2710 		node = rb_next(&__bfqq->pos_node);
2711 	else
2712 		node = rb_prev(&__bfqq->pos_node);
2713 	if (!node)
2714 		return NULL;
2715 
2716 	__bfqq = rb_entry(node, struct bfq_queue, pos_node);
2717 	if (bfq_rq_close_to_sector(__bfqq->next_rq, true, sector))
2718 		return __bfqq;
2719 
2720 	return NULL;
2721 }
2722 
2723 static struct bfq_queue *bfq_find_close_cooperator(struct bfq_data *bfqd,
2724 						   struct bfq_queue *cur_bfqq,
2725 						   sector_t sector)
2726 {
2727 	struct bfq_queue *bfqq;
2728 
2729 	/*
2730 	 * We shall notice if some of the queues are cooperating,
2731 	 * e.g., working closely on the same area of the device. In
2732 	 * that case, we can group them together and: 1) don't waste
2733 	 * time idling, and 2) serve the union of their requests in
2734 	 * the best possible order for throughput.
2735 	 */
2736 	bfqq = bfqq_find_close(bfqd, cur_bfqq, sector);
2737 	if (!bfqq || bfqq == cur_bfqq)
2738 		return NULL;
2739 
2740 	return bfqq;
2741 }
2742 
2743 static struct bfq_queue *
2744 bfq_setup_merge(struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
2745 {
2746 	int process_refs, new_process_refs;
2747 	struct bfq_queue *__bfqq;
2748 
2749 	/*
2750 	 * If there are no process references on the new_bfqq, then it is
2751 	 * unsafe to follow the ->new_bfqq chain as other bfqq's in the chain
2752 	 * may have dropped their last reference (not just their last process
2753 	 * reference).
2754 	 */
2755 	if (!bfqq_process_refs(new_bfqq))
2756 		return NULL;
2757 
2758 	/* Avoid a circular list and skip interim queue merges. */
2759 	while ((__bfqq = new_bfqq->new_bfqq)) {
2760 		if (__bfqq == bfqq)
2761 			return NULL;
2762 		new_bfqq = __bfqq;
2763 	}
2764 
2765 	process_refs = bfqq_process_refs(bfqq);
2766 	new_process_refs = bfqq_process_refs(new_bfqq);
2767 	/*
2768 	 * If the process for the bfqq has gone away, there is no
2769 	 * sense in merging the queues.
2770 	 */
2771 	if (process_refs == 0 || new_process_refs == 0)
2772 		return NULL;
2773 
2774 	/*
2775 	 * Make sure merged queues belong to the same parent. Parents could
2776 	 * have changed since the time we decided the two queues are suitable
2777 	 * for merging.
2778 	 */
2779 	if (new_bfqq->entity.parent != bfqq->entity.parent)
2780 		return NULL;
2781 
2782 	bfq_log_bfqq(bfqq->bfqd, bfqq, "scheduling merge with queue %d",
2783 		new_bfqq->pid);
2784 
2785 	/*
2786 	 * Merging is just a redirection: the requests of the process
2787 	 * owning one of the two queues are redirected to the other queue.
2788 	 * The latter queue, in its turn, is set as shared if this is the
2789 	 * first time that the requests of some process are redirected to
2790 	 * it.
2791 	 *
2792 	 * We redirect bfqq to new_bfqq and not the opposite, because
2793 	 * we are in the context of the process owning bfqq, thus we
2794 	 * have the io_cq of this process. So we can immediately
2795 	 * configure this io_cq to redirect the requests of the
2796 	 * process to new_bfqq. In contrast, the io_cq of new_bfqq is
2797 	 * not available any more (new_bfqq->bic == NULL).
2798 	 *
2799 	 * Anyway, even in case new_bfqq coincides with the in-service
2800 	 * queue, redirecting requests the in-service queue is the
2801 	 * best option, as we feed the in-service queue with new
2802 	 * requests close to the last request served and, by doing so,
2803 	 * are likely to increase the throughput.
2804 	 */
2805 	bfqq->new_bfqq = new_bfqq;
2806 	/*
2807 	 * The above assignment schedules the following redirections:
2808 	 * each time some I/O for bfqq arrives, the process that
2809 	 * generated that I/O is disassociated from bfqq and
2810 	 * associated with new_bfqq. Here we increases new_bfqq->ref
2811 	 * in advance, adding the number of processes that are
2812 	 * expected to be associated with new_bfqq as they happen to
2813 	 * issue I/O.
2814 	 */
2815 	new_bfqq->ref += process_refs;
2816 	return new_bfqq;
2817 }
2818 
2819 static bool bfq_may_be_close_cooperator(struct bfq_queue *bfqq,
2820 					struct bfq_queue *new_bfqq)
2821 {
2822 	if (bfq_too_late_for_merging(new_bfqq))
2823 		return false;
2824 
2825 	if (bfq_class_idle(bfqq) || bfq_class_idle(new_bfqq) ||
2826 	    (bfqq->ioprio_class != new_bfqq->ioprio_class))
2827 		return false;
2828 
2829 	/*
2830 	 * If either of the queues has already been detected as seeky,
2831 	 * then merging it with the other queue is unlikely to lead to
2832 	 * sequential I/O.
2833 	 */
2834 	if (BFQQ_SEEKY(bfqq) || BFQQ_SEEKY(new_bfqq))
2835 		return false;
2836 
2837 	/*
2838 	 * Interleaved I/O is known to be done by (some) applications
2839 	 * only for reads, so it does not make sense to merge async
2840 	 * queues.
2841 	 */
2842 	if (!bfq_bfqq_sync(bfqq) || !bfq_bfqq_sync(new_bfqq))
2843 		return false;
2844 
2845 	return true;
2846 }
2847 
2848 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
2849 					     struct bfq_queue *bfqq);
2850 
2851 static struct bfq_queue *
2852 bfq_setup_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2853 		       struct bfq_queue *stable_merge_bfqq,
2854 		       struct bfq_iocq_bfqq_data *bfqq_data)
2855 {
2856 	int proc_ref = min(bfqq_process_refs(bfqq),
2857 			   bfqq_process_refs(stable_merge_bfqq));
2858 	struct bfq_queue *new_bfqq = NULL;
2859 
2860 	bfqq_data->stable_merge_bfqq = NULL;
2861 	if (idling_boosts_thr_without_issues(bfqd, bfqq) || proc_ref == 0)
2862 		goto out;
2863 
2864 	/* next function will take at least one ref */
2865 	new_bfqq = bfq_setup_merge(bfqq, stable_merge_bfqq);
2866 
2867 	if (new_bfqq) {
2868 		bfqq_data->stably_merged = true;
2869 		if (new_bfqq->bic) {
2870 			unsigned int new_a_idx = new_bfqq->actuator_idx;
2871 			struct bfq_iocq_bfqq_data *new_bfqq_data =
2872 				&new_bfqq->bic->bfqq_data[new_a_idx];
2873 
2874 			new_bfqq_data->stably_merged = true;
2875 		}
2876 	}
2877 
2878 out:
2879 	/* deschedule stable merge, because done or aborted here */
2880 	bfq_put_stable_ref(stable_merge_bfqq);
2881 
2882 	return new_bfqq;
2883 }
2884 
2885 /*
2886  * Attempt to schedule a merge of bfqq with the currently in-service
2887  * queue or with a close queue among the scheduled queues.  Return
2888  * NULL if no merge was scheduled, a pointer to the shared bfq_queue
2889  * structure otherwise.
2890  *
2891  * The OOM queue is not allowed to participate to cooperation: in fact, since
2892  * the requests temporarily redirected to the OOM queue could be redirected
2893  * again to dedicated queues at any time, the state needed to correctly
2894  * handle merging with the OOM queue would be quite complex and expensive
2895  * to maintain. Besides, in such a critical condition as an out of memory,
2896  * the benefits of queue merging may be little relevant, or even negligible.
2897  *
2898  * WARNING: queue merging may impair fairness among non-weight raised
2899  * queues, for at least two reasons: 1) the original weight of a
2900  * merged queue may change during the merged state, 2) even being the
2901  * weight the same, a merged queue may be bloated with many more
2902  * requests than the ones produced by its originally-associated
2903  * process.
2904  */
2905 static struct bfq_queue *
2906 bfq_setup_cooperator(struct bfq_data *bfqd, struct bfq_queue *bfqq,
2907 		     void *io_struct, bool request, struct bfq_io_cq *bic)
2908 {
2909 	struct bfq_queue *in_service_bfqq, *new_bfqq;
2910 	unsigned int a_idx = bfqq->actuator_idx;
2911 	struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
2912 
2913 	/* if a merge has already been setup, then proceed with that first */
2914 	if (bfqq->new_bfqq)
2915 		return bfqq->new_bfqq;
2916 
2917 	/*
2918 	 * Check delayed stable merge for rotational or non-queueing
2919 	 * devs. For this branch to be executed, bfqq must not be
2920 	 * currently merged with some other queue (i.e., bfqq->bic
2921 	 * must be non null). If we considered also merged queues,
2922 	 * then we should also check whether bfqq has already been
2923 	 * merged with bic->stable_merge_bfqq. But this would be
2924 	 * costly and complicated.
2925 	 */
2926 	if (unlikely(!bfqd->nonrot_with_queueing)) {
2927 		/*
2928 		 * Make sure also that bfqq is sync, because
2929 		 * bic->stable_merge_bfqq may point to some queue (for
2930 		 * stable merging) also if bic is associated with a
2931 		 * sync queue, but this bfqq is async
2932 		 */
2933 		if (bfq_bfqq_sync(bfqq) && bfqq_data->stable_merge_bfqq &&
2934 		    !bfq_bfqq_just_created(bfqq) &&
2935 		    time_is_before_jiffies(bfqq->split_time +
2936 					  msecs_to_jiffies(bfq_late_stable_merging)) &&
2937 		    time_is_before_jiffies(bfqq->creation_time +
2938 					   msecs_to_jiffies(bfq_late_stable_merging))) {
2939 			struct bfq_queue *stable_merge_bfqq =
2940 				bfqq_data->stable_merge_bfqq;
2941 
2942 			return bfq_setup_stable_merge(bfqd, bfqq,
2943 						      stable_merge_bfqq,
2944 						      bfqq_data);
2945 		}
2946 	}
2947 
2948 	/*
2949 	 * Do not perform queue merging if the device is non
2950 	 * rotational and performs internal queueing. In fact, such a
2951 	 * device reaches a high speed through internal parallelism
2952 	 * and pipelining. This means that, to reach a high
2953 	 * throughput, it must have many requests enqueued at the same
2954 	 * time. But, in this configuration, the internal scheduling
2955 	 * algorithm of the device does exactly the job of queue
2956 	 * merging: it reorders requests so as to obtain as much as
2957 	 * possible a sequential I/O pattern. As a consequence, with
2958 	 * the workload generated by processes doing interleaved I/O,
2959 	 * the throughput reached by the device is likely to be the
2960 	 * same, with and without queue merging.
2961 	 *
2962 	 * Disabling merging also provides a remarkable benefit in
2963 	 * terms of throughput. Merging tends to make many workloads
2964 	 * artificially more uneven, because of shared queues
2965 	 * remaining non empty for incomparably more time than
2966 	 * non-merged queues. This may accentuate workload
2967 	 * asymmetries. For example, if one of the queues in a set of
2968 	 * merged queues has a higher weight than a normal queue, then
2969 	 * the shared queue may inherit such a high weight and, by
2970 	 * staying almost always active, may force BFQ to perform I/O
2971 	 * plugging most of the time. This evidently makes it harder
2972 	 * for BFQ to let the device reach a high throughput.
2973 	 *
2974 	 * Finally, the likely() macro below is not used because one
2975 	 * of the two branches is more likely than the other, but to
2976 	 * have the code path after the following if() executed as
2977 	 * fast as possible for the case of a non rotational device
2978 	 * with queueing. We want it because this is the fastest kind
2979 	 * of device. On the opposite end, the likely() may lengthen
2980 	 * the execution time of BFQ for the case of slower devices
2981 	 * (rotational or at least without queueing). But in this case
2982 	 * the execution time of BFQ matters very little, if not at
2983 	 * all.
2984 	 */
2985 	if (likely(bfqd->nonrot_with_queueing))
2986 		return NULL;
2987 
2988 	/*
2989 	 * Prevent bfqq from being merged if it has been created too
2990 	 * long ago. The idea is that true cooperating processes, and
2991 	 * thus their associated bfq_queues, are supposed to be
2992 	 * created shortly after each other. This is the case, e.g.,
2993 	 * for KVM/QEMU and dump I/O threads. Basing on this
2994 	 * assumption, the following filtering greatly reduces the
2995 	 * probability that two non-cooperating processes, which just
2996 	 * happen to do close I/O for some short time interval, have
2997 	 * their queues merged by mistake.
2998 	 */
2999 	if (bfq_too_late_for_merging(bfqq))
3000 		return NULL;
3001 
3002 	if (!io_struct || unlikely(bfqq == &bfqd->oom_bfqq))
3003 		return NULL;
3004 
3005 	/* If there is only one backlogged queue, don't search. */
3006 	if (bfq_tot_busy_queues(bfqd) == 1)
3007 		return NULL;
3008 
3009 	in_service_bfqq = bfqd->in_service_queue;
3010 
3011 	if (in_service_bfqq && in_service_bfqq != bfqq &&
3012 	    likely(in_service_bfqq != &bfqd->oom_bfqq) &&
3013 	    bfq_rq_close_to_sector(io_struct, request,
3014 				   bfqd->in_serv_last_pos) &&
3015 	    bfqq->entity.parent == in_service_bfqq->entity.parent &&
3016 	    bfq_may_be_close_cooperator(bfqq, in_service_bfqq)) {
3017 		new_bfqq = bfq_setup_merge(bfqq, in_service_bfqq);
3018 		if (new_bfqq)
3019 			return new_bfqq;
3020 	}
3021 	/*
3022 	 * Check whether there is a cooperator among currently scheduled
3023 	 * queues. The only thing we need is that the bio/request is not
3024 	 * NULL, as we need it to establish whether a cooperator exists.
3025 	 */
3026 	new_bfqq = bfq_find_close_cooperator(bfqd, bfqq,
3027 			bfq_io_struct_pos(io_struct, request));
3028 
3029 	if (new_bfqq && likely(new_bfqq != &bfqd->oom_bfqq) &&
3030 	    bfq_may_be_close_cooperator(bfqq, new_bfqq))
3031 		return bfq_setup_merge(bfqq, new_bfqq);
3032 
3033 	return NULL;
3034 }
3035 
3036 static void bfq_bfqq_save_state(struct bfq_queue *bfqq)
3037 {
3038 	struct bfq_io_cq *bic = bfqq->bic;
3039 	unsigned int a_idx = bfqq->actuator_idx;
3040 	struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[a_idx];
3041 
3042 	/*
3043 	 * If !bfqq->bic, the queue is already shared or its requests
3044 	 * have already been redirected to a shared queue; both idle window
3045 	 * and weight raising state have already been saved. Do nothing.
3046 	 */
3047 	if (!bic)
3048 		return;
3049 
3050 	bfqq_data->saved_last_serv_time_ns = bfqq->last_serv_time_ns;
3051 	bfqq_data->saved_inject_limit =	bfqq->inject_limit;
3052 	bfqq_data->saved_decrease_time_jif = bfqq->decrease_time_jif;
3053 
3054 	bfqq_data->saved_weight = bfqq->entity.orig_weight;
3055 	bfqq_data->saved_ttime = bfqq->ttime;
3056 	bfqq_data->saved_has_short_ttime =
3057 		bfq_bfqq_has_short_ttime(bfqq);
3058 	bfqq_data->saved_IO_bound = bfq_bfqq_IO_bound(bfqq);
3059 	bfqq_data->saved_io_start_time = bfqq->io_start_time;
3060 	bfqq_data->saved_tot_idle_time = bfqq->tot_idle_time;
3061 	bfqq_data->saved_in_large_burst = bfq_bfqq_in_large_burst(bfqq);
3062 	bfqq_data->was_in_burst_list =
3063 		!hlist_unhashed(&bfqq->burst_list_node);
3064 
3065 	if (unlikely(bfq_bfqq_just_created(bfqq) &&
3066 		     !bfq_bfqq_in_large_burst(bfqq) &&
3067 		     bfqq->bfqd->low_latency)) {
3068 		/*
3069 		 * bfqq being merged right after being created: bfqq
3070 		 * would have deserved interactive weight raising, but
3071 		 * did not make it to be set in a weight-raised state,
3072 		 * because of this early merge.	Store directly the
3073 		 * weight-raising state that would have been assigned
3074 		 * to bfqq, so that to avoid that bfqq unjustly fails
3075 		 * to enjoy weight raising if split soon.
3076 		 */
3077 		bfqq_data->saved_wr_coeff = bfqq->bfqd->bfq_wr_coeff;
3078 		bfqq_data->saved_wr_start_at_switch_to_srt =
3079 			bfq_smallest_from_now();
3080 		bfqq_data->saved_wr_cur_max_time =
3081 			bfq_wr_duration(bfqq->bfqd);
3082 		bfqq_data->saved_last_wr_start_finish = jiffies;
3083 	} else {
3084 		bfqq_data->saved_wr_coeff = bfqq->wr_coeff;
3085 		bfqq_data->saved_wr_start_at_switch_to_srt =
3086 			bfqq->wr_start_at_switch_to_srt;
3087 		bfqq_data->saved_service_from_wr =
3088 			bfqq->service_from_wr;
3089 		bfqq_data->saved_last_wr_start_finish =
3090 			bfqq->last_wr_start_finish;
3091 		bfqq_data->saved_wr_cur_max_time = bfqq->wr_cur_max_time;
3092 	}
3093 }
3094 
3095 
3096 static void
3097 bfq_reassign_last_bfqq(struct bfq_queue *cur_bfqq, struct bfq_queue *new_bfqq)
3098 {
3099 	if (cur_bfqq->entity.parent &&
3100 	    cur_bfqq->entity.parent->last_bfqq_created == cur_bfqq)
3101 		cur_bfqq->entity.parent->last_bfqq_created = new_bfqq;
3102 	else if (cur_bfqq->bfqd && cur_bfqq->bfqd->last_bfqq_created == cur_bfqq)
3103 		cur_bfqq->bfqd->last_bfqq_created = new_bfqq;
3104 }
3105 
3106 void bfq_release_process_ref(struct bfq_data *bfqd, struct bfq_queue *bfqq)
3107 {
3108 	/*
3109 	 * To prevent bfqq's service guarantees from being violated,
3110 	 * bfqq may be left busy, i.e., queued for service, even if
3111 	 * empty (see comments in __bfq_bfqq_expire() for
3112 	 * details). But, if no process will send requests to bfqq any
3113 	 * longer, then there is no point in keeping bfqq queued for
3114 	 * service. In addition, keeping bfqq queued for service, but
3115 	 * with no process ref any longer, may have caused bfqq to be
3116 	 * freed when dequeued from service. But this is assumed to
3117 	 * never happen.
3118 	 */
3119 	if (bfq_bfqq_busy(bfqq) && RB_EMPTY_ROOT(&bfqq->sort_list) &&
3120 	    bfqq != bfqd->in_service_queue)
3121 		bfq_del_bfqq_busy(bfqq, false);
3122 
3123 	bfq_reassign_last_bfqq(bfqq, NULL);
3124 
3125 	bfq_put_queue(bfqq);
3126 }
3127 
3128 static void
3129 bfq_merge_bfqqs(struct bfq_data *bfqd, struct bfq_io_cq *bic,
3130 		struct bfq_queue *bfqq, struct bfq_queue *new_bfqq)
3131 {
3132 	bfq_log_bfqq(bfqd, bfqq, "merging with queue %lu",
3133 		(unsigned long)new_bfqq->pid);
3134 	/* Save weight raising and idle window of the merged queues */
3135 	bfq_bfqq_save_state(bfqq);
3136 	bfq_bfqq_save_state(new_bfqq);
3137 	if (bfq_bfqq_IO_bound(bfqq))
3138 		bfq_mark_bfqq_IO_bound(new_bfqq);
3139 	bfq_clear_bfqq_IO_bound(bfqq);
3140 
3141 	/*
3142 	 * The processes associated with bfqq are cooperators of the
3143 	 * processes associated with new_bfqq. So, if bfqq has a
3144 	 * waker, then assume that all these processes will be happy
3145 	 * to let bfqq's waker freely inject I/O when they have no
3146 	 * I/O.
3147 	 */
3148 	if (bfqq->waker_bfqq && !new_bfqq->waker_bfqq &&
3149 	    bfqq->waker_bfqq != new_bfqq) {
3150 		new_bfqq->waker_bfqq = bfqq->waker_bfqq;
3151 		new_bfqq->tentative_waker_bfqq = NULL;
3152 
3153 		/*
3154 		 * If the waker queue disappears, then
3155 		 * new_bfqq->waker_bfqq must be reset. So insert
3156 		 * new_bfqq into the woken_list of the waker. See
3157 		 * bfq_check_waker for details.
3158 		 */
3159 		hlist_add_head(&new_bfqq->woken_list_node,
3160 			       &new_bfqq->waker_bfqq->woken_list);
3161 
3162 	}
3163 
3164 	/*
3165 	 * If bfqq is weight-raised, then let new_bfqq inherit
3166 	 * weight-raising. To reduce false positives, neglect the case
3167 	 * where bfqq has just been created, but has not yet made it
3168 	 * to be weight-raised (which may happen because EQM may merge
3169 	 * bfqq even before bfq_add_request is executed for the first
3170 	 * time for bfqq). Handling this case would however be very
3171 	 * easy, thanks to the flag just_created.
3172 	 */
3173 	if (new_bfqq->wr_coeff == 1 && bfqq->wr_coeff > 1) {
3174 		new_bfqq->wr_coeff = bfqq->wr_coeff;
3175 		new_bfqq->wr_cur_max_time = bfqq->wr_cur_max_time;
3176 		new_bfqq->last_wr_start_finish = bfqq->last_wr_start_finish;
3177 		new_bfqq->wr_start_at_switch_to_srt =
3178 			bfqq->wr_start_at_switch_to_srt;
3179 		if (bfq_bfqq_busy(new_bfqq))
3180 			bfqd->wr_busy_queues++;
3181 		new_bfqq->entity.prio_changed = 1;
3182 	}
3183 
3184 	if (bfqq->wr_coeff > 1) { /* bfqq has given its wr to new_bfqq */
3185 		bfqq->wr_coeff = 1;
3186 		bfqq->entity.prio_changed = 1;
3187 		if (bfq_bfqq_busy(bfqq))
3188 			bfqd->wr_busy_queues--;
3189 	}
3190 
3191 	bfq_log_bfqq(bfqd, new_bfqq, "merge_bfqqs: wr_busy %d",
3192 		     bfqd->wr_busy_queues);
3193 
3194 	/*
3195 	 * Merge queues (that is, let bic redirect its requests to new_bfqq)
3196 	 */
3197 	bic_set_bfqq(bic, new_bfqq, true, bfqq->actuator_idx);
3198 	bfq_mark_bfqq_coop(new_bfqq);
3199 	/*
3200 	 * new_bfqq now belongs to at least two bics (it is a shared queue):
3201 	 * set new_bfqq->bic to NULL. bfqq either:
3202 	 * - does not belong to any bic any more, and hence bfqq->bic must
3203 	 *   be set to NULL, or
3204 	 * - is a queue whose owning bics have already been redirected to a
3205 	 *   different queue, hence the queue is destined to not belong to
3206 	 *   any bic soon and bfqq->bic is already NULL (therefore the next
3207 	 *   assignment causes no harm).
3208 	 */
3209 	new_bfqq->bic = NULL;
3210 	/*
3211 	 * If the queue is shared, the pid is the pid of one of the associated
3212 	 * processes. Which pid depends on the exact sequence of merge events
3213 	 * the queue underwent. So printing such a pid is useless and confusing
3214 	 * because it reports a random pid between those of the associated
3215 	 * processes.
3216 	 * We mark such a queue with a pid -1, and then print SHARED instead of
3217 	 * a pid in logging messages.
3218 	 */
3219 	new_bfqq->pid = -1;
3220 	bfqq->bic = NULL;
3221 
3222 	bfq_reassign_last_bfqq(bfqq, new_bfqq);
3223 
3224 	bfq_release_process_ref(bfqd, bfqq);
3225 }
3226 
3227 static bool bfq_allow_bio_merge(struct request_queue *q, struct request *rq,
3228 				struct bio *bio)
3229 {
3230 	struct bfq_data *bfqd = q->elevator->elevator_data;
3231 	bool is_sync = op_is_sync(bio->bi_opf);
3232 	struct bfq_queue *bfqq = bfqd->bio_bfqq, *new_bfqq;
3233 
3234 	/*
3235 	 * Disallow merge of a sync bio into an async request.
3236 	 */
3237 	if (is_sync && !rq_is_sync(rq))
3238 		return false;
3239 
3240 	/*
3241 	 * Lookup the bfqq that this bio will be queued with. Allow
3242 	 * merge only if rq is queued there.
3243 	 */
3244 	if (!bfqq)
3245 		return false;
3246 
3247 	/*
3248 	 * We take advantage of this function to perform an early merge
3249 	 * of the queues of possible cooperating processes.
3250 	 */
3251 	new_bfqq = bfq_setup_cooperator(bfqd, bfqq, bio, false, bfqd->bio_bic);
3252 	if (new_bfqq) {
3253 		/*
3254 		 * bic still points to bfqq, then it has not yet been
3255 		 * redirected to some other bfq_queue, and a queue
3256 		 * merge between bfqq and new_bfqq can be safely
3257 		 * fulfilled, i.e., bic can be redirected to new_bfqq
3258 		 * and bfqq can be put.
3259 		 */
3260 		bfq_merge_bfqqs(bfqd, bfqd->bio_bic, bfqq,
3261 				new_bfqq);
3262 		/*
3263 		 * If we get here, bio will be queued into new_queue,
3264 		 * so use new_bfqq to decide whether bio and rq can be
3265 		 * merged.
3266 		 */
3267 		bfqq = new_bfqq;
3268 
3269 		/*
3270 		 * Change also bqfd->bio_bfqq, as
3271 		 * bfqd->bio_bic now points to new_bfqq, and
3272 		 * this function may be invoked again (and then may
3273 		 * use again bqfd->bio_bfqq).
3274 		 */
3275 		bfqd->bio_bfqq = bfqq;
3276 	}
3277 
3278 	return bfqq == RQ_BFQQ(rq);
3279 }
3280 
3281 /*
3282  * Set the maximum time for the in-service queue to consume its
3283  * budget. This prevents seeky processes from lowering the throughput.
3284  * In practice, a time-slice service scheme is used with seeky
3285  * processes.
3286  */
3287 static void bfq_set_budget_timeout(struct bfq_data *bfqd,
3288 				   struct bfq_queue *bfqq)
3289 {
3290 	unsigned int timeout_coeff;
3291 
3292 	if (bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time)
3293 		timeout_coeff = 1;
3294 	else
3295 		timeout_coeff = bfqq->entity.weight / bfqq->entity.orig_weight;
3296 
3297 	bfqd->last_budget_start = ktime_get();
3298 
3299 	bfqq->budget_timeout = jiffies +
3300 		bfqd->bfq_timeout * timeout_coeff;
3301 }
3302 
3303 static void __bfq_set_in_service_queue(struct bfq_data *bfqd,
3304 				       struct bfq_queue *bfqq)
3305 {
3306 	if (bfqq) {
3307 		bfq_clear_bfqq_fifo_expire(bfqq);
3308 
3309 		bfqd->budgets_assigned = (bfqd->budgets_assigned * 7 + 256) / 8;
3310 
3311 		if (time_is_before_jiffies(bfqq->last_wr_start_finish) &&
3312 		    bfqq->wr_coeff > 1 &&
3313 		    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
3314 		    time_is_before_jiffies(bfqq->budget_timeout)) {
3315 			/*
3316 			 * For soft real-time queues, move the start
3317 			 * of the weight-raising period forward by the
3318 			 * time the queue has not received any
3319 			 * service. Otherwise, a relatively long
3320 			 * service delay is likely to cause the
3321 			 * weight-raising period of the queue to end,
3322 			 * because of the short duration of the
3323 			 * weight-raising period of a soft real-time
3324 			 * queue.  It is worth noting that this move
3325 			 * is not so dangerous for the other queues,
3326 			 * because soft real-time queues are not
3327 			 * greedy.
3328 			 *
3329 			 * To not add a further variable, we use the
3330 			 * overloaded field budget_timeout to
3331 			 * determine for how long the queue has not
3332 			 * received service, i.e., how much time has
3333 			 * elapsed since the queue expired. However,
3334 			 * this is a little imprecise, because
3335 			 * budget_timeout is set to jiffies if bfqq
3336 			 * not only expires, but also remains with no
3337 			 * request.
3338 			 */
3339 			if (time_after(bfqq->budget_timeout,
3340 				       bfqq->last_wr_start_finish))
3341 				bfqq->last_wr_start_finish +=
3342 					jiffies - bfqq->budget_timeout;
3343 			else
3344 				bfqq->last_wr_start_finish = jiffies;
3345 		}
3346 
3347 		bfq_set_budget_timeout(bfqd, bfqq);
3348 		bfq_log_bfqq(bfqd, bfqq,
3349 			     "set_in_service_queue, cur-budget = %d",
3350 			     bfqq->entity.budget);
3351 	}
3352 
3353 	bfqd->in_service_queue = bfqq;
3354 	bfqd->in_serv_last_pos = 0;
3355 }
3356 
3357 /*
3358  * Get and set a new queue for service.
3359  */
3360 static struct bfq_queue *bfq_set_in_service_queue(struct bfq_data *bfqd)
3361 {
3362 	struct bfq_queue *bfqq = bfq_get_next_queue(bfqd);
3363 
3364 	__bfq_set_in_service_queue(bfqd, bfqq);
3365 	return bfqq;
3366 }
3367 
3368 static void bfq_arm_slice_timer(struct bfq_data *bfqd)
3369 {
3370 	struct bfq_queue *bfqq = bfqd->in_service_queue;
3371 	u32 sl;
3372 
3373 	bfq_mark_bfqq_wait_request(bfqq);
3374 
3375 	/*
3376 	 * We don't want to idle for seeks, but we do want to allow
3377 	 * fair distribution of slice time for a process doing back-to-back
3378 	 * seeks. So allow a little bit of time for him to submit a new rq.
3379 	 */
3380 	sl = bfqd->bfq_slice_idle;
3381 	/*
3382 	 * Unless the queue is being weight-raised or the scenario is
3383 	 * asymmetric, grant only minimum idle time if the queue
3384 	 * is seeky. A long idling is preserved for a weight-raised
3385 	 * queue, or, more in general, in an asymmetric scenario,
3386 	 * because a long idling is needed for guaranteeing to a queue
3387 	 * its reserved share of the throughput (in particular, it is
3388 	 * needed if the queue has a higher weight than some other
3389 	 * queue).
3390 	 */
3391 	if (BFQQ_SEEKY(bfqq) && bfqq->wr_coeff == 1 &&
3392 	    !bfq_asymmetric_scenario(bfqd, bfqq))
3393 		sl = min_t(u64, sl, BFQ_MIN_TT);
3394 	else if (bfqq->wr_coeff > 1)
3395 		sl = max_t(u32, sl, 20ULL * NSEC_PER_MSEC);
3396 
3397 	bfqd->last_idling_start = ktime_get();
3398 	bfqd->last_idling_start_jiffies = jiffies;
3399 
3400 	hrtimer_start(&bfqd->idle_slice_timer, ns_to_ktime(sl),
3401 		      HRTIMER_MODE_REL);
3402 	bfqg_stats_set_start_idle_time(bfqq_group(bfqq));
3403 }
3404 
3405 /*
3406  * In autotuning mode, max_budget is dynamically recomputed as the
3407  * amount of sectors transferred in timeout at the estimated peak
3408  * rate. This enables BFQ to utilize a full timeslice with a full
3409  * budget, even if the in-service queue is served at peak rate. And
3410  * this maximises throughput with sequential workloads.
3411  */
3412 static unsigned long bfq_calc_max_budget(struct bfq_data *bfqd)
3413 {
3414 	return (u64)bfqd->peak_rate * USEC_PER_MSEC *
3415 		jiffies_to_msecs(bfqd->bfq_timeout)>>BFQ_RATE_SHIFT;
3416 }
3417 
3418 /*
3419  * Update parameters related to throughput and responsiveness, as a
3420  * function of the estimated peak rate. See comments on
3421  * bfq_calc_max_budget(), and on the ref_wr_duration array.
3422  */
3423 static void update_thr_responsiveness_params(struct bfq_data *bfqd)
3424 {
3425 	if (bfqd->bfq_user_max_budget == 0) {
3426 		bfqd->bfq_max_budget =
3427 			bfq_calc_max_budget(bfqd);
3428 		bfq_log(bfqd, "new max_budget = %d", bfqd->bfq_max_budget);
3429 	}
3430 }
3431 
3432 static void bfq_reset_rate_computation(struct bfq_data *bfqd,
3433 				       struct request *rq)
3434 {
3435 	if (rq != NULL) { /* new rq dispatch now, reset accordingly */
3436 		bfqd->last_dispatch = bfqd->first_dispatch = ktime_get_ns();
3437 		bfqd->peak_rate_samples = 1;
3438 		bfqd->sequential_samples = 0;
3439 		bfqd->tot_sectors_dispatched = bfqd->last_rq_max_size =
3440 			blk_rq_sectors(rq);
3441 	} else /* no new rq dispatched, just reset the number of samples */
3442 		bfqd->peak_rate_samples = 0; /* full re-init on next disp. */
3443 
3444 	bfq_log(bfqd,
3445 		"reset_rate_computation at end, sample %u/%u tot_sects %llu",
3446 		bfqd->peak_rate_samples, bfqd->sequential_samples,
3447 		bfqd->tot_sectors_dispatched);
3448 }
3449 
3450 static void bfq_update_rate_reset(struct bfq_data *bfqd, struct request *rq)
3451 {
3452 	u32 rate, weight, divisor;
3453 
3454 	/*
3455 	 * For the convergence property to hold (see comments on
3456 	 * bfq_update_peak_rate()) and for the assessment to be
3457 	 * reliable, a minimum number of samples must be present, and
3458 	 * a minimum amount of time must have elapsed. If not so, do
3459 	 * not compute new rate. Just reset parameters, to get ready
3460 	 * for a new evaluation attempt.
3461 	 */
3462 	if (bfqd->peak_rate_samples < BFQ_RATE_MIN_SAMPLES ||
3463 	    bfqd->delta_from_first < BFQ_RATE_MIN_INTERVAL)
3464 		goto reset_computation;
3465 
3466 	/*
3467 	 * If a new request completion has occurred after last
3468 	 * dispatch, then, to approximate the rate at which requests
3469 	 * have been served by the device, it is more precise to
3470 	 * extend the observation interval to the last completion.
3471 	 */
3472 	bfqd->delta_from_first =
3473 		max_t(u64, bfqd->delta_from_first,
3474 		      bfqd->last_completion - bfqd->first_dispatch);
3475 
3476 	/*
3477 	 * Rate computed in sects/usec, and not sects/nsec, for
3478 	 * precision issues.
3479 	 */
3480 	rate = div64_ul(bfqd->tot_sectors_dispatched<<BFQ_RATE_SHIFT,
3481 			div_u64(bfqd->delta_from_first, NSEC_PER_USEC));
3482 
3483 	/*
3484 	 * Peak rate not updated if:
3485 	 * - the percentage of sequential dispatches is below 3/4 of the
3486 	 *   total, and rate is below the current estimated peak rate
3487 	 * - rate is unreasonably high (> 20M sectors/sec)
3488 	 */
3489 	if ((bfqd->sequential_samples < (3 * bfqd->peak_rate_samples)>>2 &&
3490 	     rate <= bfqd->peak_rate) ||
3491 		rate > 20<<BFQ_RATE_SHIFT)
3492 		goto reset_computation;
3493 
3494 	/*
3495 	 * We have to update the peak rate, at last! To this purpose,
3496 	 * we use a low-pass filter. We compute the smoothing constant
3497 	 * of the filter as a function of the 'weight' of the new
3498 	 * measured rate.
3499 	 *
3500 	 * As can be seen in next formulas, we define this weight as a
3501 	 * quantity proportional to how sequential the workload is,
3502 	 * and to how long the observation time interval is.
3503 	 *
3504 	 * The weight runs from 0 to 8. The maximum value of the
3505 	 * weight, 8, yields the minimum value for the smoothing
3506 	 * constant. At this minimum value for the smoothing constant,
3507 	 * the measured rate contributes for half of the next value of
3508 	 * the estimated peak rate.
3509 	 *
3510 	 * So, the first step is to compute the weight as a function
3511 	 * of how sequential the workload is. Note that the weight
3512 	 * cannot reach 9, because bfqd->sequential_samples cannot
3513 	 * become equal to bfqd->peak_rate_samples, which, in its
3514 	 * turn, holds true because bfqd->sequential_samples is not
3515 	 * incremented for the first sample.
3516 	 */
3517 	weight = (9 * bfqd->sequential_samples) / bfqd->peak_rate_samples;
3518 
3519 	/*
3520 	 * Second step: further refine the weight as a function of the
3521 	 * duration of the observation interval.
3522 	 */
3523 	weight = min_t(u32, 8,
3524 		       div_u64(weight * bfqd->delta_from_first,
3525 			       BFQ_RATE_REF_INTERVAL));
3526 
3527 	/*
3528 	 * Divisor ranging from 10, for minimum weight, to 2, for
3529 	 * maximum weight.
3530 	 */
3531 	divisor = 10 - weight;
3532 
3533 	/*
3534 	 * Finally, update peak rate:
3535 	 *
3536 	 * peak_rate = peak_rate * (divisor-1) / divisor  +  rate / divisor
3537 	 */
3538 	bfqd->peak_rate *= divisor-1;
3539 	bfqd->peak_rate /= divisor;
3540 	rate /= divisor; /* smoothing constant alpha = 1/divisor */
3541 
3542 	bfqd->peak_rate += rate;
3543 
3544 	/*
3545 	 * For a very slow device, bfqd->peak_rate can reach 0 (see
3546 	 * the minimum representable values reported in the comments
3547 	 * on BFQ_RATE_SHIFT). Push to 1 if this happens, to avoid
3548 	 * divisions by zero where bfqd->peak_rate is used as a
3549 	 * divisor.
3550 	 */
3551 	bfqd->peak_rate = max_t(u32, 1, bfqd->peak_rate);
3552 
3553 	update_thr_responsiveness_params(bfqd);
3554 
3555 reset_computation:
3556 	bfq_reset_rate_computation(bfqd, rq);
3557 }
3558 
3559 /*
3560  * Update the read/write peak rate (the main quantity used for
3561  * auto-tuning, see update_thr_responsiveness_params()).
3562  *
3563  * It is not trivial to estimate the peak rate (correctly): because of
3564  * the presence of sw and hw queues between the scheduler and the
3565  * device components that finally serve I/O requests, it is hard to
3566  * say exactly when a given dispatched request is served inside the
3567  * device, and for how long. As a consequence, it is hard to know
3568  * precisely at what rate a given set of requests is actually served
3569  * by the device.
3570  *
3571  * On the opposite end, the dispatch time of any request is trivially
3572  * available, and, from this piece of information, the "dispatch rate"
3573  * of requests can be immediately computed. So, the idea in the next
3574  * function is to use what is known, namely request dispatch times
3575  * (plus, when useful, request completion times), to estimate what is
3576  * unknown, namely in-device request service rate.
3577  *
3578  * The main issue is that, because of the above facts, the rate at
3579  * which a certain set of requests is dispatched over a certain time
3580  * interval can vary greatly with respect to the rate at which the
3581  * same requests are then served. But, since the size of any
3582  * intermediate queue is limited, and the service scheme is lossless
3583  * (no request is silently dropped), the following obvious convergence
3584  * property holds: the number of requests dispatched MUST become
3585  * closer and closer to the number of requests completed as the
3586  * observation interval grows. This is the key property used in
3587  * the next function to estimate the peak service rate as a function
3588  * of the observed dispatch rate. The function assumes to be invoked
3589  * on every request dispatch.
3590  */
3591 static void bfq_update_peak_rate(struct bfq_data *bfqd, struct request *rq)
3592 {
3593 	u64 now_ns = ktime_get_ns();
3594 
3595 	if (bfqd->peak_rate_samples == 0) { /* first dispatch */
3596 		bfq_log(bfqd, "update_peak_rate: goto reset, samples %d",
3597 			bfqd->peak_rate_samples);
3598 		bfq_reset_rate_computation(bfqd, rq);
3599 		goto update_last_values; /* will add one sample */
3600 	}
3601 
3602 	/*
3603 	 * Device idle for very long: the observation interval lasting
3604 	 * up to this dispatch cannot be a valid observation interval
3605 	 * for computing a new peak rate (similarly to the late-
3606 	 * completion event in bfq_completed_request()). Go to
3607 	 * update_rate_and_reset to have the following three steps
3608 	 * taken:
3609 	 * - close the observation interval at the last (previous)
3610 	 *   request dispatch or completion
3611 	 * - compute rate, if possible, for that observation interval
3612 	 * - start a new observation interval with this dispatch
3613 	 */
3614 	if (now_ns - bfqd->last_dispatch > 100*NSEC_PER_MSEC &&
3615 	    bfqd->tot_rq_in_driver == 0)
3616 		goto update_rate_and_reset;
3617 
3618 	/* Update sampling information */
3619 	bfqd->peak_rate_samples++;
3620 
3621 	if ((bfqd->tot_rq_in_driver > 0 ||
3622 		now_ns - bfqd->last_completion < BFQ_MIN_TT)
3623 	    && !BFQ_RQ_SEEKY(bfqd, bfqd->last_position, rq))
3624 		bfqd->sequential_samples++;
3625 
3626 	bfqd->tot_sectors_dispatched += blk_rq_sectors(rq);
3627 
3628 	/* Reset max observed rq size every 32 dispatches */
3629 	if (likely(bfqd->peak_rate_samples % 32))
3630 		bfqd->last_rq_max_size =
3631 			max_t(u32, blk_rq_sectors(rq), bfqd->last_rq_max_size);
3632 	else
3633 		bfqd->last_rq_max_size = blk_rq_sectors(rq);
3634 
3635 	bfqd->delta_from_first = now_ns - bfqd->first_dispatch;
3636 
3637 	/* Target observation interval not yet reached, go on sampling */
3638 	if (bfqd->delta_from_first < BFQ_RATE_REF_INTERVAL)
3639 		goto update_last_values;
3640 
3641 update_rate_and_reset:
3642 	bfq_update_rate_reset(bfqd, rq);
3643 update_last_values:
3644 	bfqd->last_position = blk_rq_pos(rq) + blk_rq_sectors(rq);
3645 	if (RQ_BFQQ(rq) == bfqd->in_service_queue)
3646 		bfqd->in_serv_last_pos = bfqd->last_position;
3647 	bfqd->last_dispatch = now_ns;
3648 }
3649 
3650 /*
3651  * Remove request from internal lists.
3652  */
3653 static void bfq_dispatch_remove(struct request_queue *q, struct request *rq)
3654 {
3655 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
3656 
3657 	/*
3658 	 * For consistency, the next instruction should have been
3659 	 * executed after removing the request from the queue and
3660 	 * dispatching it.  We execute instead this instruction before
3661 	 * bfq_remove_request() (and hence introduce a temporary
3662 	 * inconsistency), for efficiency.  In fact, should this
3663 	 * dispatch occur for a non in-service bfqq, this anticipated
3664 	 * increment prevents two counters related to bfqq->dispatched
3665 	 * from risking to be, first, uselessly decremented, and then
3666 	 * incremented again when the (new) value of bfqq->dispatched
3667 	 * happens to be taken into account.
3668 	 */
3669 	bfqq->dispatched++;
3670 	bfq_update_peak_rate(q->elevator->elevator_data, rq);
3671 
3672 	bfq_remove_request(q, rq);
3673 }
3674 
3675 /*
3676  * There is a case where idling does not have to be performed for
3677  * throughput concerns, but to preserve the throughput share of
3678  * the process associated with bfqq.
3679  *
3680  * To introduce this case, we can note that allowing the drive
3681  * to enqueue more than one request at a time, and hence
3682  * delegating de facto final scheduling decisions to the
3683  * drive's internal scheduler, entails loss of control on the
3684  * actual request service order. In particular, the critical
3685  * situation is when requests from different processes happen
3686  * to be present, at the same time, in the internal queue(s)
3687  * of the drive. In such a situation, the drive, by deciding
3688  * the service order of the internally-queued requests, does
3689  * determine also the actual throughput distribution among
3690  * these processes. But the drive typically has no notion or
3691  * concern about per-process throughput distribution, and
3692  * makes its decisions only on a per-request basis. Therefore,
3693  * the service distribution enforced by the drive's internal
3694  * scheduler is likely to coincide with the desired throughput
3695  * distribution only in a completely symmetric, or favorably
3696  * skewed scenario where:
3697  * (i-a) each of these processes must get the same throughput as
3698  *	 the others,
3699  * (i-b) in case (i-a) does not hold, it holds that the process
3700  *       associated with bfqq must receive a lower or equal
3701  *	 throughput than any of the other processes;
3702  * (ii)  the I/O of each process has the same properties, in
3703  *       terms of locality (sequential or random), direction
3704  *       (reads or writes), request sizes, greediness
3705  *       (from I/O-bound to sporadic), and so on;
3706 
3707  * In fact, in such a scenario, the drive tends to treat the requests
3708  * of each process in about the same way as the requests of the
3709  * others, and thus to provide each of these processes with about the
3710  * same throughput.  This is exactly the desired throughput
3711  * distribution if (i-a) holds, or, if (i-b) holds instead, this is an
3712  * even more convenient distribution for (the process associated with)
3713  * bfqq.
3714  *
3715  * In contrast, in any asymmetric or unfavorable scenario, device
3716  * idling (I/O-dispatch plugging) is certainly needed to guarantee
3717  * that bfqq receives its assigned fraction of the device throughput
3718  * (see [1] for details).
3719  *
3720  * The problem is that idling may significantly reduce throughput with
3721  * certain combinations of types of I/O and devices. An important
3722  * example is sync random I/O on flash storage with command
3723  * queueing. So, unless bfqq falls in cases where idling also boosts
3724  * throughput, it is important to check conditions (i-a), i(-b) and
3725  * (ii) accurately, so as to avoid idling when not strictly needed for
3726  * service guarantees.
3727  *
3728  * Unfortunately, it is extremely difficult to thoroughly check
3729  * condition (ii). And, in case there are active groups, it becomes
3730  * very difficult to check conditions (i-a) and (i-b) too.  In fact,
3731  * if there are active groups, then, for conditions (i-a) or (i-b) to
3732  * become false 'indirectly', it is enough that an active group
3733  * contains more active processes or sub-groups than some other active
3734  * group. More precisely, for conditions (i-a) or (i-b) to become
3735  * false because of such a group, it is not even necessary that the
3736  * group is (still) active: it is sufficient that, even if the group
3737  * has become inactive, some of its descendant processes still have
3738  * some request already dispatched but still waiting for
3739  * completion. In fact, requests have still to be guaranteed their
3740  * share of the throughput even after being dispatched. In this
3741  * respect, it is easy to show that, if a group frequently becomes
3742  * inactive while still having in-flight requests, and if, when this
3743  * happens, the group is not considered in the calculation of whether
3744  * the scenario is asymmetric, then the group may fail to be
3745  * guaranteed its fair share of the throughput (basically because
3746  * idling may not be performed for the descendant processes of the
3747  * group, but it had to be).  We address this issue with the following
3748  * bi-modal behavior, implemented in the function
3749  * bfq_asymmetric_scenario().
3750  *
3751  * If there are groups with requests waiting for completion
3752  * (as commented above, some of these groups may even be
3753  * already inactive), then the scenario is tagged as
3754  * asymmetric, conservatively, without checking any of the
3755  * conditions (i-a), (i-b) or (ii). So the device is idled for bfqq.
3756  * This behavior matches also the fact that groups are created
3757  * exactly if controlling I/O is a primary concern (to
3758  * preserve bandwidth and latency guarantees).
3759  *
3760  * On the opposite end, if there are no groups with requests waiting
3761  * for completion, then only conditions (i-a) and (i-b) are actually
3762  * controlled, i.e., provided that conditions (i-a) or (i-b) holds,
3763  * idling is not performed, regardless of whether condition (ii)
3764  * holds.  In other words, only if conditions (i-a) and (i-b) do not
3765  * hold, then idling is allowed, and the device tends to be prevented
3766  * from queueing many requests, possibly of several processes. Since
3767  * there are no groups with requests waiting for completion, then, to
3768  * control conditions (i-a) and (i-b) it is enough to check just
3769  * whether all the queues with requests waiting for completion also
3770  * have the same weight.
3771  *
3772  * Not checking condition (ii) evidently exposes bfqq to the
3773  * risk of getting less throughput than its fair share.
3774  * However, for queues with the same weight, a further
3775  * mechanism, preemption, mitigates or even eliminates this
3776  * problem. And it does so without consequences on overall
3777  * throughput. This mechanism and its benefits are explained
3778  * in the next three paragraphs.
3779  *
3780  * Even if a queue, say Q, is expired when it remains idle, Q
3781  * can still preempt the new in-service queue if the next
3782  * request of Q arrives soon (see the comments on
3783  * bfq_bfqq_update_budg_for_activation). If all queues and
3784  * groups have the same weight, this form of preemption,
3785  * combined with the hole-recovery heuristic described in the
3786  * comments on function bfq_bfqq_update_budg_for_activation,
3787  * are enough to preserve a correct bandwidth distribution in
3788  * the mid term, even without idling. In fact, even if not
3789  * idling allows the internal queues of the device to contain
3790  * many requests, and thus to reorder requests, we can rather
3791  * safely assume that the internal scheduler still preserves a
3792  * minimum of mid-term fairness.
3793  *
3794  * More precisely, this preemption-based, idleless approach
3795  * provides fairness in terms of IOPS, and not sectors per
3796  * second. This can be seen with a simple example. Suppose
3797  * that there are two queues with the same weight, but that
3798  * the first queue receives requests of 8 sectors, while the
3799  * second queue receives requests of 1024 sectors. In
3800  * addition, suppose that each of the two queues contains at
3801  * most one request at a time, which implies that each queue
3802  * always remains idle after it is served. Finally, after
3803  * remaining idle, each queue receives very quickly a new
3804  * request. It follows that the two queues are served
3805  * alternatively, preempting each other if needed. This
3806  * implies that, although both queues have the same weight,
3807  * the queue with large requests receives a service that is
3808  * 1024/8 times as high as the service received by the other
3809  * queue.
3810  *
3811  * The motivation for using preemption instead of idling (for
3812  * queues with the same weight) is that, by not idling,
3813  * service guarantees are preserved (completely or at least in
3814  * part) without minimally sacrificing throughput. And, if
3815  * there is no active group, then the primary expectation for
3816  * this device is probably a high throughput.
3817  *
3818  * We are now left only with explaining the two sub-conditions in the
3819  * additional compound condition that is checked below for deciding
3820  * whether the scenario is asymmetric. To explain the first
3821  * sub-condition, we need to add that the function
3822  * bfq_asymmetric_scenario checks the weights of only
3823  * non-weight-raised queues, for efficiency reasons (see comments on
3824  * bfq_weights_tree_add()). Then the fact that bfqq is weight-raised
3825  * is checked explicitly here. More precisely, the compound condition
3826  * below takes into account also the fact that, even if bfqq is being
3827  * weight-raised, the scenario is still symmetric if all queues with
3828  * requests waiting for completion happen to be
3829  * weight-raised. Actually, we should be even more precise here, and
3830  * differentiate between interactive weight raising and soft real-time
3831  * weight raising.
3832  *
3833  * The second sub-condition checked in the compound condition is
3834  * whether there is a fair amount of already in-flight I/O not
3835  * belonging to bfqq. If so, I/O dispatching is to be plugged, for the
3836  * following reason. The drive may decide to serve in-flight
3837  * non-bfqq's I/O requests before bfqq's ones, thereby delaying the
3838  * arrival of new I/O requests for bfqq (recall that bfqq is sync). If
3839  * I/O-dispatching is not plugged, then, while bfqq remains empty, a
3840  * basically uncontrolled amount of I/O from other queues may be
3841  * dispatched too, possibly causing the service of bfqq's I/O to be
3842  * delayed even longer in the drive. This problem gets more and more
3843  * serious as the speed and the queue depth of the drive grow,
3844  * because, as these two quantities grow, the probability to find no
3845  * queue busy but many requests in flight grows too. By contrast,
3846  * plugging I/O dispatching minimizes the delay induced by already
3847  * in-flight I/O, and enables bfqq to recover the bandwidth it may
3848  * lose because of this delay.
3849  *
3850  * As a side note, it is worth considering that the above
3851  * device-idling countermeasures may however fail in the following
3852  * unlucky scenario: if I/O-dispatch plugging is (correctly) disabled
3853  * in a time period during which all symmetry sub-conditions hold, and
3854  * therefore the device is allowed to enqueue many requests, but at
3855  * some later point in time some sub-condition stops to hold, then it
3856  * may become impossible to make requests be served in the desired
3857  * order until all the requests already queued in the device have been
3858  * served. The last sub-condition commented above somewhat mitigates
3859  * this problem for weight-raised queues.
3860  *
3861  * However, as an additional mitigation for this problem, we preserve
3862  * plugging for a special symmetric case that may suddenly turn into
3863  * asymmetric: the case where only bfqq is busy. In this case, not
3864  * expiring bfqq does not cause any harm to any other queues in terms
3865  * of service guarantees. In contrast, it avoids the following unlucky
3866  * sequence of events: (1) bfqq is expired, (2) a new queue with a
3867  * lower weight than bfqq becomes busy (or more queues), (3) the new
3868  * queue is served until a new request arrives for bfqq, (4) when bfqq
3869  * is finally served, there are so many requests of the new queue in
3870  * the drive that the pending requests for bfqq take a lot of time to
3871  * be served. In particular, event (2) may case even already
3872  * dispatched requests of bfqq to be delayed, inside the drive. So, to
3873  * avoid this series of events, the scenario is preventively declared
3874  * as asymmetric also if bfqq is the only busy queues
3875  */
3876 static bool idling_needed_for_service_guarantees(struct bfq_data *bfqd,
3877 						 struct bfq_queue *bfqq)
3878 {
3879 	int tot_busy_queues = bfq_tot_busy_queues(bfqd);
3880 
3881 	/* No point in idling for bfqq if it won't get requests any longer */
3882 	if (unlikely(!bfqq_process_refs(bfqq)))
3883 		return false;
3884 
3885 	return (bfqq->wr_coeff > 1 &&
3886 		(bfqd->wr_busy_queues < tot_busy_queues ||
3887 		 bfqd->tot_rq_in_driver >= bfqq->dispatched + 4)) ||
3888 		bfq_asymmetric_scenario(bfqd, bfqq) ||
3889 		tot_busy_queues == 1;
3890 }
3891 
3892 static bool __bfq_bfqq_expire(struct bfq_data *bfqd, struct bfq_queue *bfqq,
3893 			      enum bfqq_expiration reason)
3894 {
3895 	/*
3896 	 * If this bfqq is shared between multiple processes, check
3897 	 * to make sure that those processes are still issuing I/Os
3898 	 * within the mean seek distance. If not, it may be time to
3899 	 * break the queues apart again.
3900 	 */
3901 	if (bfq_bfqq_coop(bfqq) && BFQQ_SEEKY(bfqq))
3902 		bfq_mark_bfqq_split_coop(bfqq);
3903 
3904 	/*
3905 	 * Consider queues with a higher finish virtual time than
3906 	 * bfqq. If idling_needed_for_service_guarantees(bfqq) returns
3907 	 * true, then bfqq's bandwidth would be violated if an
3908 	 * uncontrolled amount of I/O from these queues were
3909 	 * dispatched while bfqq is waiting for its new I/O to
3910 	 * arrive. This is exactly what may happen if this is a forced
3911 	 * expiration caused by a preemption attempt, and if bfqq is
3912 	 * not re-scheduled. To prevent this from happening, re-queue
3913 	 * bfqq if it needs I/O-dispatch plugging, even if it is
3914 	 * empty. By doing so, bfqq is granted to be served before the
3915 	 * above queues (provided that bfqq is of course eligible).
3916 	 */
3917 	if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
3918 	    !(reason == BFQQE_PREEMPTED &&
3919 	      idling_needed_for_service_guarantees(bfqd, bfqq))) {
3920 		if (bfqq->dispatched == 0)
3921 			/*
3922 			 * Overloading budget_timeout field to store
3923 			 * the time at which the queue remains with no
3924 			 * backlog and no outstanding request; used by
3925 			 * the weight-raising mechanism.
3926 			 */
3927 			bfqq->budget_timeout = jiffies;
3928 
3929 		bfq_del_bfqq_busy(bfqq, true);
3930 	} else {
3931 		bfq_requeue_bfqq(bfqd, bfqq, true);
3932 		/*
3933 		 * Resort priority tree of potential close cooperators.
3934 		 * See comments on bfq_pos_tree_add_move() for the unlikely().
3935 		 */
3936 		if (unlikely(!bfqd->nonrot_with_queueing &&
3937 			     !RB_EMPTY_ROOT(&bfqq->sort_list)))
3938 			bfq_pos_tree_add_move(bfqd, bfqq);
3939 	}
3940 
3941 	/*
3942 	 * All in-service entities must have been properly deactivated
3943 	 * or requeued before executing the next function, which
3944 	 * resets all in-service entities as no more in service. This
3945 	 * may cause bfqq to be freed. If this happens, the next
3946 	 * function returns true.
3947 	 */
3948 	return __bfq_bfqd_reset_in_service(bfqd);
3949 }
3950 
3951 /**
3952  * __bfq_bfqq_recalc_budget - try to adapt the budget to the @bfqq behavior.
3953  * @bfqd: device data.
3954  * @bfqq: queue to update.
3955  * @reason: reason for expiration.
3956  *
3957  * Handle the feedback on @bfqq budget at queue expiration.
3958  * See the body for detailed comments.
3959  */
3960 static void __bfq_bfqq_recalc_budget(struct bfq_data *bfqd,
3961 				     struct bfq_queue *bfqq,
3962 				     enum bfqq_expiration reason)
3963 {
3964 	struct request *next_rq;
3965 	int budget, min_budget;
3966 
3967 	min_budget = bfq_min_budget(bfqd);
3968 
3969 	if (bfqq->wr_coeff == 1)
3970 		budget = bfqq->max_budget;
3971 	else /*
3972 	      * Use a constant, low budget for weight-raised queues,
3973 	      * to help achieve a low latency. Keep it slightly higher
3974 	      * than the minimum possible budget, to cause a little
3975 	      * bit fewer expirations.
3976 	      */
3977 		budget = 2 * min_budget;
3978 
3979 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last budg %d, budg left %d",
3980 		bfqq->entity.budget, bfq_bfqq_budget_left(bfqq));
3981 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: last max_budg %d, min budg %d",
3982 		budget, bfq_min_budget(bfqd));
3983 	bfq_log_bfqq(bfqd, bfqq, "recalc_budg: sync %d, seeky %d",
3984 		bfq_bfqq_sync(bfqq), BFQQ_SEEKY(bfqd->in_service_queue));
3985 
3986 	if (bfq_bfqq_sync(bfqq) && bfqq->wr_coeff == 1) {
3987 		switch (reason) {
3988 		/*
3989 		 * Caveat: in all the following cases we trade latency
3990 		 * for throughput.
3991 		 */
3992 		case BFQQE_TOO_IDLE:
3993 			/*
3994 			 * This is the only case where we may reduce
3995 			 * the budget: if there is no request of the
3996 			 * process still waiting for completion, then
3997 			 * we assume (tentatively) that the timer has
3998 			 * expired because the batch of requests of
3999 			 * the process could have been served with a
4000 			 * smaller budget.  Hence, betting that
4001 			 * process will behave in the same way when it
4002 			 * becomes backlogged again, we reduce its
4003 			 * next budget.  As long as we guess right,
4004 			 * this budget cut reduces the latency
4005 			 * experienced by the process.
4006 			 *
4007 			 * However, if there are still outstanding
4008 			 * requests, then the process may have not yet
4009 			 * issued its next request just because it is
4010 			 * still waiting for the completion of some of
4011 			 * the still outstanding ones.  So in this
4012 			 * subcase we do not reduce its budget, on the
4013 			 * contrary we increase it to possibly boost
4014 			 * the throughput, as discussed in the
4015 			 * comments to the BUDGET_TIMEOUT case.
4016 			 */
4017 			if (bfqq->dispatched > 0) /* still outstanding reqs */
4018 				budget = min(budget * 2, bfqd->bfq_max_budget);
4019 			else {
4020 				if (budget > 5 * min_budget)
4021 					budget -= 4 * min_budget;
4022 				else
4023 					budget = min_budget;
4024 			}
4025 			break;
4026 		case BFQQE_BUDGET_TIMEOUT:
4027 			/*
4028 			 * We double the budget here because it gives
4029 			 * the chance to boost the throughput if this
4030 			 * is not a seeky process (and has bumped into
4031 			 * this timeout because of, e.g., ZBR).
4032 			 */
4033 			budget = min(budget * 2, bfqd->bfq_max_budget);
4034 			break;
4035 		case BFQQE_BUDGET_EXHAUSTED:
4036 			/*
4037 			 * The process still has backlog, and did not
4038 			 * let either the budget timeout or the disk
4039 			 * idling timeout expire. Hence it is not
4040 			 * seeky, has a short thinktime and may be
4041 			 * happy with a higher budget too. So
4042 			 * definitely increase the budget of this good
4043 			 * candidate to boost the disk throughput.
4044 			 */
4045 			budget = min(budget * 4, bfqd->bfq_max_budget);
4046 			break;
4047 		case BFQQE_NO_MORE_REQUESTS:
4048 			/*
4049 			 * For queues that expire for this reason, it
4050 			 * is particularly important to keep the
4051 			 * budget close to the actual service they
4052 			 * need. Doing so reduces the timestamp
4053 			 * misalignment problem described in the
4054 			 * comments in the body of
4055 			 * __bfq_activate_entity. In fact, suppose
4056 			 * that a queue systematically expires for
4057 			 * BFQQE_NO_MORE_REQUESTS and presents a
4058 			 * new request in time to enjoy timestamp
4059 			 * back-shifting. The larger the budget of the
4060 			 * queue is with respect to the service the
4061 			 * queue actually requests in each service
4062 			 * slot, the more times the queue can be
4063 			 * reactivated with the same virtual finish
4064 			 * time. It follows that, even if this finish
4065 			 * time is pushed to the system virtual time
4066 			 * to reduce the consequent timestamp
4067 			 * misalignment, the queue unjustly enjoys for
4068 			 * many re-activations a lower finish time
4069 			 * than all newly activated queues.
4070 			 *
4071 			 * The service needed by bfqq is measured
4072 			 * quite precisely by bfqq->entity.service.
4073 			 * Since bfqq does not enjoy device idling,
4074 			 * bfqq->entity.service is equal to the number
4075 			 * of sectors that the process associated with
4076 			 * bfqq requested to read/write before waiting
4077 			 * for request completions, or blocking for
4078 			 * other reasons.
4079 			 */
4080 			budget = max_t(int, bfqq->entity.service, min_budget);
4081 			break;
4082 		default:
4083 			return;
4084 		}
4085 	} else if (!bfq_bfqq_sync(bfqq)) {
4086 		/*
4087 		 * Async queues get always the maximum possible
4088 		 * budget, as for them we do not care about latency
4089 		 * (in addition, their ability to dispatch is limited
4090 		 * by the charging factor).
4091 		 */
4092 		budget = bfqd->bfq_max_budget;
4093 	}
4094 
4095 	bfqq->max_budget = budget;
4096 
4097 	if (bfqd->budgets_assigned >= bfq_stats_min_budgets &&
4098 	    !bfqd->bfq_user_max_budget)
4099 		bfqq->max_budget = min(bfqq->max_budget, bfqd->bfq_max_budget);
4100 
4101 	/*
4102 	 * If there is still backlog, then assign a new budget, making
4103 	 * sure that it is large enough for the next request.  Since
4104 	 * the finish time of bfqq must be kept in sync with the
4105 	 * budget, be sure to call __bfq_bfqq_expire() *after* this
4106 	 * update.
4107 	 *
4108 	 * If there is no backlog, then no need to update the budget;
4109 	 * it will be updated on the arrival of a new request.
4110 	 */
4111 	next_rq = bfqq->next_rq;
4112 	if (next_rq)
4113 		bfqq->entity.budget = max_t(unsigned long, bfqq->max_budget,
4114 					    bfq_serv_to_charge(next_rq, bfqq));
4115 
4116 	bfq_log_bfqq(bfqd, bfqq, "head sect: %u, new budget %d",
4117 			next_rq ? blk_rq_sectors(next_rq) : 0,
4118 			bfqq->entity.budget);
4119 }
4120 
4121 /*
4122  * Return true if the process associated with bfqq is "slow". The slow
4123  * flag is used, in addition to the budget timeout, to reduce the
4124  * amount of service provided to seeky processes, and thus reduce
4125  * their chances to lower the throughput. More details in the comments
4126  * on the function bfq_bfqq_expire().
4127  *
4128  * An important observation is in order: as discussed in the comments
4129  * on the function bfq_update_peak_rate(), with devices with internal
4130  * queues, it is hard if ever possible to know when and for how long
4131  * an I/O request is processed by the device (apart from the trivial
4132  * I/O pattern where a new request is dispatched only after the
4133  * previous one has been completed). This makes it hard to evaluate
4134  * the real rate at which the I/O requests of each bfq_queue are
4135  * served.  In fact, for an I/O scheduler like BFQ, serving a
4136  * bfq_queue means just dispatching its requests during its service
4137  * slot (i.e., until the budget of the queue is exhausted, or the
4138  * queue remains idle, or, finally, a timeout fires). But, during the
4139  * service slot of a bfq_queue, around 100 ms at most, the device may
4140  * be even still processing requests of bfq_queues served in previous
4141  * service slots. On the opposite end, the requests of the in-service
4142  * bfq_queue may be completed after the service slot of the queue
4143  * finishes.
4144  *
4145  * Anyway, unless more sophisticated solutions are used
4146  * (where possible), the sum of the sizes of the requests dispatched
4147  * during the service slot of a bfq_queue is probably the only
4148  * approximation available for the service received by the bfq_queue
4149  * during its service slot. And this sum is the quantity used in this
4150  * function to evaluate the I/O speed of a process.
4151  */
4152 static bool bfq_bfqq_is_slow(struct bfq_data *bfqd, struct bfq_queue *bfqq,
4153 				 bool compensate, unsigned long *delta_ms)
4154 {
4155 	ktime_t delta_ktime;
4156 	u32 delta_usecs;
4157 	bool slow = BFQQ_SEEKY(bfqq); /* if delta too short, use seekyness */
4158 
4159 	if (!bfq_bfqq_sync(bfqq))
4160 		return false;
4161 
4162 	if (compensate)
4163 		delta_ktime = bfqd->last_idling_start;
4164 	else
4165 		delta_ktime = ktime_get();
4166 	delta_ktime = ktime_sub(delta_ktime, bfqd->last_budget_start);
4167 	delta_usecs = ktime_to_us(delta_ktime);
4168 
4169 	/* don't use too short time intervals */
4170 	if (delta_usecs < 1000) {
4171 		if (blk_queue_nonrot(bfqd->queue))
4172 			 /*
4173 			  * give same worst-case guarantees as idling
4174 			  * for seeky
4175 			  */
4176 			*delta_ms = BFQ_MIN_TT / NSEC_PER_MSEC;
4177 		else /* charge at least one seek */
4178 			*delta_ms = bfq_slice_idle / NSEC_PER_MSEC;
4179 
4180 		return slow;
4181 	}
4182 
4183 	*delta_ms = delta_usecs / USEC_PER_MSEC;
4184 
4185 	/*
4186 	 * Use only long (> 20ms) intervals to filter out excessive
4187 	 * spikes in service rate estimation.
4188 	 */
4189 	if (delta_usecs > 20000) {
4190 		/*
4191 		 * Caveat for rotational devices: processes doing I/O
4192 		 * in the slower disk zones tend to be slow(er) even
4193 		 * if not seeky. In this respect, the estimated peak
4194 		 * rate is likely to be an average over the disk
4195 		 * surface. Accordingly, to not be too harsh with
4196 		 * unlucky processes, a process is deemed slow only if
4197 		 * its rate has been lower than half of the estimated
4198 		 * peak rate.
4199 		 */
4200 		slow = bfqq->entity.service < bfqd->bfq_max_budget / 2;
4201 	}
4202 
4203 	bfq_log_bfqq(bfqd, bfqq, "bfq_bfqq_is_slow: slow %d", slow);
4204 
4205 	return slow;
4206 }
4207 
4208 /*
4209  * To be deemed as soft real-time, an application must meet two
4210  * requirements. First, the application must not require an average
4211  * bandwidth higher than the approximate bandwidth required to playback or
4212  * record a compressed high-definition video.
4213  * The next function is invoked on the completion of the last request of a
4214  * batch, to compute the next-start time instant, soft_rt_next_start, such
4215  * that, if the next request of the application does not arrive before
4216  * soft_rt_next_start, then the above requirement on the bandwidth is met.
4217  *
4218  * The second requirement is that the request pattern of the application is
4219  * isochronous, i.e., that, after issuing a request or a batch of requests,
4220  * the application stops issuing new requests until all its pending requests
4221  * have been completed. After that, the application may issue a new batch,
4222  * and so on.
4223  * For this reason the next function is invoked to compute
4224  * soft_rt_next_start only for applications that meet this requirement,
4225  * whereas soft_rt_next_start is set to infinity for applications that do
4226  * not.
4227  *
4228  * Unfortunately, even a greedy (i.e., I/O-bound) application may
4229  * happen to meet, occasionally or systematically, both the above
4230  * bandwidth and isochrony requirements. This may happen at least in
4231  * the following circumstances. First, if the CPU load is high. The
4232  * application may stop issuing requests while the CPUs are busy
4233  * serving other processes, then restart, then stop again for a while,
4234  * and so on. The other circumstances are related to the storage
4235  * device: the storage device is highly loaded or reaches a low-enough
4236  * throughput with the I/O of the application (e.g., because the I/O
4237  * is random and/or the device is slow). In all these cases, the
4238  * I/O of the application may be simply slowed down enough to meet
4239  * the bandwidth and isochrony requirements. To reduce the probability
4240  * that greedy applications are deemed as soft real-time in these
4241  * corner cases, a further rule is used in the computation of
4242  * soft_rt_next_start: the return value of this function is forced to
4243  * be higher than the maximum between the following two quantities.
4244  *
4245  * (a) Current time plus: (1) the maximum time for which the arrival
4246  *     of a request is waited for when a sync queue becomes idle,
4247  *     namely bfqd->bfq_slice_idle, and (2) a few extra jiffies. We
4248  *     postpone for a moment the reason for adding a few extra
4249  *     jiffies; we get back to it after next item (b).  Lower-bounding
4250  *     the return value of this function with the current time plus
4251  *     bfqd->bfq_slice_idle tends to filter out greedy applications,
4252  *     because the latter issue their next request as soon as possible
4253  *     after the last one has been completed. In contrast, a soft
4254  *     real-time application spends some time processing data, after a
4255  *     batch of its requests has been completed.
4256  *
4257  * (b) Current value of bfqq->soft_rt_next_start. As pointed out
4258  *     above, greedy applications may happen to meet both the
4259  *     bandwidth and isochrony requirements under heavy CPU or
4260  *     storage-device load. In more detail, in these scenarios, these
4261  *     applications happen, only for limited time periods, to do I/O
4262  *     slowly enough to meet all the requirements described so far,
4263  *     including the filtering in above item (a). These slow-speed
4264  *     time intervals are usually interspersed between other time
4265  *     intervals during which these applications do I/O at a very high
4266  *     speed. Fortunately, exactly because of the high speed of the
4267  *     I/O in the high-speed intervals, the values returned by this
4268  *     function happen to be so high, near the end of any such
4269  *     high-speed interval, to be likely to fall *after* the end of
4270  *     the low-speed time interval that follows. These high values are
4271  *     stored in bfqq->soft_rt_next_start after each invocation of
4272  *     this function. As a consequence, if the last value of
4273  *     bfqq->soft_rt_next_start is constantly used to lower-bound the
4274  *     next value that this function may return, then, from the very
4275  *     beginning of a low-speed interval, bfqq->soft_rt_next_start is
4276  *     likely to be constantly kept so high that any I/O request
4277  *     issued during the low-speed interval is considered as arriving
4278  *     to soon for the application to be deemed as soft
4279  *     real-time. Then, in the high-speed interval that follows, the
4280  *     application will not be deemed as soft real-time, just because
4281  *     it will do I/O at a high speed. And so on.
4282  *
4283  * Getting back to the filtering in item (a), in the following two
4284  * cases this filtering might be easily passed by a greedy
4285  * application, if the reference quantity was just
4286  * bfqd->bfq_slice_idle:
4287  * 1) HZ is so low that the duration of a jiffy is comparable to or
4288  *    higher than bfqd->bfq_slice_idle. This happens, e.g., on slow
4289  *    devices with HZ=100. The time granularity may be so coarse
4290  *    that the approximation, in jiffies, of bfqd->bfq_slice_idle
4291  *    is rather lower than the exact value.
4292  * 2) jiffies, instead of increasing at a constant rate, may stop increasing
4293  *    for a while, then suddenly 'jump' by several units to recover the lost
4294  *    increments. This seems to happen, e.g., inside virtual machines.
4295  * To address this issue, in the filtering in (a) we do not use as a
4296  * reference time interval just bfqd->bfq_slice_idle, but
4297  * bfqd->bfq_slice_idle plus a few jiffies. In particular, we add the
4298  * minimum number of jiffies for which the filter seems to be quite
4299  * precise also in embedded systems and KVM/QEMU virtual machines.
4300  */
4301 static unsigned long bfq_bfqq_softrt_next_start(struct bfq_data *bfqd,
4302 						struct bfq_queue *bfqq)
4303 {
4304 	return max3(bfqq->soft_rt_next_start,
4305 		    bfqq->last_idle_bklogged +
4306 		    HZ * bfqq->service_from_backlogged /
4307 		    bfqd->bfq_wr_max_softrt_rate,
4308 		    jiffies + nsecs_to_jiffies(bfqq->bfqd->bfq_slice_idle) + 4);
4309 }
4310 
4311 /**
4312  * bfq_bfqq_expire - expire a queue.
4313  * @bfqd: device owning the queue.
4314  * @bfqq: the queue to expire.
4315  * @compensate: if true, compensate for the time spent idling.
4316  * @reason: the reason causing the expiration.
4317  *
4318  * If the process associated with bfqq does slow I/O (e.g., because it
4319  * issues random requests), we charge bfqq with the time it has been
4320  * in service instead of the service it has received (see
4321  * bfq_bfqq_charge_time for details on how this goal is achieved). As
4322  * a consequence, bfqq will typically get higher timestamps upon
4323  * reactivation, and hence it will be rescheduled as if it had
4324  * received more service than what it has actually received. In the
4325  * end, bfqq receives less service in proportion to how slowly its
4326  * associated process consumes its budgets (and hence how seriously it
4327  * tends to lower the throughput). In addition, this time-charging
4328  * strategy guarantees time fairness among slow processes. In
4329  * contrast, if the process associated with bfqq is not slow, we
4330  * charge bfqq exactly with the service it has received.
4331  *
4332  * Charging time to the first type of queues and the exact service to
4333  * the other has the effect of using the WF2Q+ policy to schedule the
4334  * former on a timeslice basis, without violating service domain
4335  * guarantees among the latter.
4336  */
4337 void bfq_bfqq_expire(struct bfq_data *bfqd,
4338 		     struct bfq_queue *bfqq,
4339 		     bool compensate,
4340 		     enum bfqq_expiration reason)
4341 {
4342 	bool slow;
4343 	unsigned long delta = 0;
4344 	struct bfq_entity *entity = &bfqq->entity;
4345 
4346 	/*
4347 	 * Check whether the process is slow (see bfq_bfqq_is_slow).
4348 	 */
4349 	slow = bfq_bfqq_is_slow(bfqd, bfqq, compensate, &delta);
4350 
4351 	/*
4352 	 * As above explained, charge slow (typically seeky) and
4353 	 * timed-out queues with the time and not the service
4354 	 * received, to favor sequential workloads.
4355 	 *
4356 	 * Processes doing I/O in the slower disk zones will tend to
4357 	 * be slow(er) even if not seeky. Therefore, since the
4358 	 * estimated peak rate is actually an average over the disk
4359 	 * surface, these processes may timeout just for bad luck. To
4360 	 * avoid punishing them, do not charge time to processes that
4361 	 * succeeded in consuming at least 2/3 of their budget. This
4362 	 * allows BFQ to preserve enough elasticity to still perform
4363 	 * bandwidth, and not time, distribution with little unlucky
4364 	 * or quasi-sequential processes.
4365 	 */
4366 	if (bfqq->wr_coeff == 1 &&
4367 	    (slow ||
4368 	     (reason == BFQQE_BUDGET_TIMEOUT &&
4369 	      bfq_bfqq_budget_left(bfqq) >=  entity->budget / 3)))
4370 		bfq_bfqq_charge_time(bfqd, bfqq, delta);
4371 
4372 	if (bfqd->low_latency && bfqq->wr_coeff == 1)
4373 		bfqq->last_wr_start_finish = jiffies;
4374 
4375 	if (bfqd->low_latency && bfqd->bfq_wr_max_softrt_rate > 0 &&
4376 	    RB_EMPTY_ROOT(&bfqq->sort_list)) {
4377 		/*
4378 		 * If we get here, and there are no outstanding
4379 		 * requests, then the request pattern is isochronous
4380 		 * (see the comments on the function
4381 		 * bfq_bfqq_softrt_next_start()). Therefore we can
4382 		 * compute soft_rt_next_start.
4383 		 *
4384 		 * If, instead, the queue still has outstanding
4385 		 * requests, then we have to wait for the completion
4386 		 * of all the outstanding requests to discover whether
4387 		 * the request pattern is actually isochronous.
4388 		 */
4389 		if (bfqq->dispatched == 0)
4390 			bfqq->soft_rt_next_start =
4391 				bfq_bfqq_softrt_next_start(bfqd, bfqq);
4392 		else if (bfqq->dispatched > 0) {
4393 			/*
4394 			 * Schedule an update of soft_rt_next_start to when
4395 			 * the task may be discovered to be isochronous.
4396 			 */
4397 			bfq_mark_bfqq_softrt_update(bfqq);
4398 		}
4399 	}
4400 
4401 	bfq_log_bfqq(bfqd, bfqq,
4402 		"expire (%d, slow %d, num_disp %d, short_ttime %d)", reason,
4403 		slow, bfqq->dispatched, bfq_bfqq_has_short_ttime(bfqq));
4404 
4405 	/*
4406 	 * bfqq expired, so no total service time needs to be computed
4407 	 * any longer: reset state machine for measuring total service
4408 	 * times.
4409 	 */
4410 	bfqd->rqs_injected = bfqd->wait_dispatch = false;
4411 	bfqd->waited_rq = NULL;
4412 
4413 	/*
4414 	 * Increase, decrease or leave budget unchanged according to
4415 	 * reason.
4416 	 */
4417 	__bfq_bfqq_recalc_budget(bfqd, bfqq, reason);
4418 	if (__bfq_bfqq_expire(bfqd, bfqq, reason))
4419 		/* bfqq is gone, no more actions on it */
4420 		return;
4421 
4422 	/* mark bfqq as waiting a request only if a bic still points to it */
4423 	if (!bfq_bfqq_busy(bfqq) &&
4424 	    reason != BFQQE_BUDGET_TIMEOUT &&
4425 	    reason != BFQQE_BUDGET_EXHAUSTED) {
4426 		bfq_mark_bfqq_non_blocking_wait_rq(bfqq);
4427 		/*
4428 		 * Not setting service to 0, because, if the next rq
4429 		 * arrives in time, the queue will go on receiving
4430 		 * service with this same budget (as if it never expired)
4431 		 */
4432 	} else
4433 		entity->service = 0;
4434 
4435 	/*
4436 	 * Reset the received-service counter for every parent entity.
4437 	 * Differently from what happens with bfqq->entity.service,
4438 	 * the resetting of this counter never needs to be postponed
4439 	 * for parent entities. In fact, in case bfqq may have a
4440 	 * chance to go on being served using the last, partially
4441 	 * consumed budget, bfqq->entity.service needs to be kept,
4442 	 * because if bfqq then actually goes on being served using
4443 	 * the same budget, the last value of bfqq->entity.service is
4444 	 * needed to properly decrement bfqq->entity.budget by the
4445 	 * portion already consumed. In contrast, it is not necessary
4446 	 * to keep entity->service for parent entities too, because
4447 	 * the bubble up of the new value of bfqq->entity.budget will
4448 	 * make sure that the budgets of parent entities are correct,
4449 	 * even in case bfqq and thus parent entities go on receiving
4450 	 * service with the same budget.
4451 	 */
4452 	entity = entity->parent;
4453 	for_each_entity(entity)
4454 		entity->service = 0;
4455 }
4456 
4457 /*
4458  * Budget timeout is not implemented through a dedicated timer, but
4459  * just checked on request arrivals and completions, as well as on
4460  * idle timer expirations.
4461  */
4462 static bool bfq_bfqq_budget_timeout(struct bfq_queue *bfqq)
4463 {
4464 	return time_is_before_eq_jiffies(bfqq->budget_timeout);
4465 }
4466 
4467 /*
4468  * If we expire a queue that is actively waiting (i.e., with the
4469  * device idled) for the arrival of a new request, then we may incur
4470  * the timestamp misalignment problem described in the body of the
4471  * function __bfq_activate_entity. Hence we return true only if this
4472  * condition does not hold, or if the queue is slow enough to deserve
4473  * only to be kicked off for preserving a high throughput.
4474  */
4475 static bool bfq_may_expire_for_budg_timeout(struct bfq_queue *bfqq)
4476 {
4477 	bfq_log_bfqq(bfqq->bfqd, bfqq,
4478 		"may_budget_timeout: wait_request %d left %d timeout %d",
4479 		bfq_bfqq_wait_request(bfqq),
4480 			bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3,
4481 		bfq_bfqq_budget_timeout(bfqq));
4482 
4483 	return (!bfq_bfqq_wait_request(bfqq) ||
4484 		bfq_bfqq_budget_left(bfqq) >=  bfqq->entity.budget / 3)
4485 		&&
4486 		bfq_bfqq_budget_timeout(bfqq);
4487 }
4488 
4489 static bool idling_boosts_thr_without_issues(struct bfq_data *bfqd,
4490 					     struct bfq_queue *bfqq)
4491 {
4492 	bool rot_without_queueing =
4493 		!blk_queue_nonrot(bfqd->queue) && !bfqd->hw_tag,
4494 		bfqq_sequential_and_IO_bound,
4495 		idling_boosts_thr;
4496 
4497 	/* No point in idling for bfqq if it won't get requests any longer */
4498 	if (unlikely(!bfqq_process_refs(bfqq)))
4499 		return false;
4500 
4501 	bfqq_sequential_and_IO_bound = !BFQQ_SEEKY(bfqq) &&
4502 		bfq_bfqq_IO_bound(bfqq) && bfq_bfqq_has_short_ttime(bfqq);
4503 
4504 	/*
4505 	 * The next variable takes into account the cases where idling
4506 	 * boosts the throughput.
4507 	 *
4508 	 * The value of the variable is computed considering, first, that
4509 	 * idling is virtually always beneficial for the throughput if:
4510 	 * (a) the device is not NCQ-capable and rotational, or
4511 	 * (b) regardless of the presence of NCQ, the device is rotational and
4512 	 *     the request pattern for bfqq is I/O-bound and sequential, or
4513 	 * (c) regardless of whether it is rotational, the device is
4514 	 *     not NCQ-capable and the request pattern for bfqq is
4515 	 *     I/O-bound and sequential.
4516 	 *
4517 	 * Secondly, and in contrast to the above item (b), idling an
4518 	 * NCQ-capable flash-based device would not boost the
4519 	 * throughput even with sequential I/O; rather it would lower
4520 	 * the throughput in proportion to how fast the device
4521 	 * is. Accordingly, the next variable is true if any of the
4522 	 * above conditions (a), (b) or (c) is true, and, in
4523 	 * particular, happens to be false if bfqd is an NCQ-capable
4524 	 * flash-based device.
4525 	 */
4526 	idling_boosts_thr = rot_without_queueing ||
4527 		((!blk_queue_nonrot(bfqd->queue) || !bfqd->hw_tag) &&
4528 		 bfqq_sequential_and_IO_bound);
4529 
4530 	/*
4531 	 * The return value of this function is equal to that of
4532 	 * idling_boosts_thr, unless a special case holds. In this
4533 	 * special case, described below, idling may cause problems to
4534 	 * weight-raised queues.
4535 	 *
4536 	 * When the request pool is saturated (e.g., in the presence
4537 	 * of write hogs), if the processes associated with
4538 	 * non-weight-raised queues ask for requests at a lower rate,
4539 	 * then processes associated with weight-raised queues have a
4540 	 * higher probability to get a request from the pool
4541 	 * immediately (or at least soon) when they need one. Thus
4542 	 * they have a higher probability to actually get a fraction
4543 	 * of the device throughput proportional to their high
4544 	 * weight. This is especially true with NCQ-capable drives,
4545 	 * which enqueue several requests in advance, and further
4546 	 * reorder internally-queued requests.
4547 	 *
4548 	 * For this reason, we force to false the return value if
4549 	 * there are weight-raised busy queues. In this case, and if
4550 	 * bfqq is not weight-raised, this guarantees that the device
4551 	 * is not idled for bfqq (if, instead, bfqq is weight-raised,
4552 	 * then idling will be guaranteed by another variable, see
4553 	 * below). Combined with the timestamping rules of BFQ (see
4554 	 * [1] for details), this behavior causes bfqq, and hence any
4555 	 * sync non-weight-raised queue, to get a lower number of
4556 	 * requests served, and thus to ask for a lower number of
4557 	 * requests from the request pool, before the busy
4558 	 * weight-raised queues get served again. This often mitigates
4559 	 * starvation problems in the presence of heavy write
4560 	 * workloads and NCQ, thereby guaranteeing a higher
4561 	 * application and system responsiveness in these hostile
4562 	 * scenarios.
4563 	 */
4564 	return idling_boosts_thr &&
4565 		bfqd->wr_busy_queues == 0;
4566 }
4567 
4568 /*
4569  * For a queue that becomes empty, device idling is allowed only if
4570  * this function returns true for that queue. As a consequence, since
4571  * device idling plays a critical role for both throughput boosting
4572  * and service guarantees, the return value of this function plays a
4573  * critical role as well.
4574  *
4575  * In a nutshell, this function returns true only if idling is
4576  * beneficial for throughput or, even if detrimental for throughput,
4577  * idling is however necessary to preserve service guarantees (low
4578  * latency, desired throughput distribution, ...). In particular, on
4579  * NCQ-capable devices, this function tries to return false, so as to
4580  * help keep the drives' internal queues full, whenever this helps the
4581  * device boost the throughput without causing any service-guarantee
4582  * issue.
4583  *
4584  * Most of the issues taken into account to get the return value of
4585  * this function are not trivial. We discuss these issues in the two
4586  * functions providing the main pieces of information needed by this
4587  * function.
4588  */
4589 static bool bfq_better_to_idle(struct bfq_queue *bfqq)
4590 {
4591 	struct bfq_data *bfqd = bfqq->bfqd;
4592 	bool idling_boosts_thr_with_no_issue, idling_needed_for_service_guar;
4593 
4594 	/* No point in idling for bfqq if it won't get requests any longer */
4595 	if (unlikely(!bfqq_process_refs(bfqq)))
4596 		return false;
4597 
4598 	if (unlikely(bfqd->strict_guarantees))
4599 		return true;
4600 
4601 	/*
4602 	 * Idling is performed only if slice_idle > 0. In addition, we
4603 	 * do not idle if
4604 	 * (a) bfqq is async
4605 	 * (b) bfqq is in the idle io prio class: in this case we do
4606 	 * not idle because we want to minimize the bandwidth that
4607 	 * queues in this class can steal to higher-priority queues
4608 	 */
4609 	if (bfqd->bfq_slice_idle == 0 || !bfq_bfqq_sync(bfqq) ||
4610 	   bfq_class_idle(bfqq))
4611 		return false;
4612 
4613 	idling_boosts_thr_with_no_issue =
4614 		idling_boosts_thr_without_issues(bfqd, bfqq);
4615 
4616 	idling_needed_for_service_guar =
4617 		idling_needed_for_service_guarantees(bfqd, bfqq);
4618 
4619 	/*
4620 	 * We have now the two components we need to compute the
4621 	 * return value of the function, which is true only if idling
4622 	 * either boosts the throughput (without issues), or is
4623 	 * necessary to preserve service guarantees.
4624 	 */
4625 	return idling_boosts_thr_with_no_issue ||
4626 		idling_needed_for_service_guar;
4627 }
4628 
4629 /*
4630  * If the in-service queue is empty but the function bfq_better_to_idle
4631  * returns true, then:
4632  * 1) the queue must remain in service and cannot be expired, and
4633  * 2) the device must be idled to wait for the possible arrival of a new
4634  *    request for the queue.
4635  * See the comments on the function bfq_better_to_idle for the reasons
4636  * why performing device idling is the best choice to boost the throughput
4637  * and preserve service guarantees when bfq_better_to_idle itself
4638  * returns true.
4639  */
4640 static bool bfq_bfqq_must_idle(struct bfq_queue *bfqq)
4641 {
4642 	return RB_EMPTY_ROOT(&bfqq->sort_list) && bfq_better_to_idle(bfqq);
4643 }
4644 
4645 /*
4646  * This function chooses the queue from which to pick the next extra
4647  * I/O request to inject, if it finds a compatible queue. See the
4648  * comments on bfq_update_inject_limit() for details on the injection
4649  * mechanism, and for the definitions of the quantities mentioned
4650  * below.
4651  */
4652 static struct bfq_queue *
4653 bfq_choose_bfqq_for_injection(struct bfq_data *bfqd)
4654 {
4655 	struct bfq_queue *bfqq, *in_serv_bfqq = bfqd->in_service_queue;
4656 	unsigned int limit = in_serv_bfqq->inject_limit;
4657 	int i;
4658 
4659 	/*
4660 	 * If
4661 	 * - bfqq is not weight-raised and therefore does not carry
4662 	 *   time-critical I/O,
4663 	 * or
4664 	 * - regardless of whether bfqq is weight-raised, bfqq has
4665 	 *   however a long think time, during which it can absorb the
4666 	 *   effect of an appropriate number of extra I/O requests
4667 	 *   from other queues (see bfq_update_inject_limit for
4668 	 *   details on the computation of this number);
4669 	 * then injection can be performed without restrictions.
4670 	 */
4671 	bool in_serv_always_inject = in_serv_bfqq->wr_coeff == 1 ||
4672 		!bfq_bfqq_has_short_ttime(in_serv_bfqq);
4673 
4674 	/*
4675 	 * If
4676 	 * - the baseline total service time could not be sampled yet,
4677 	 *   so the inject limit happens to be still 0, and
4678 	 * - a lot of time has elapsed since the plugging of I/O
4679 	 *   dispatching started, so drive speed is being wasted
4680 	 *   significantly;
4681 	 * then temporarily raise inject limit to one request.
4682 	 */
4683 	if (limit == 0 && in_serv_bfqq->last_serv_time_ns == 0 &&
4684 	    bfq_bfqq_wait_request(in_serv_bfqq) &&
4685 	    time_is_before_eq_jiffies(bfqd->last_idling_start_jiffies +
4686 				      bfqd->bfq_slice_idle)
4687 		)
4688 		limit = 1;
4689 
4690 	if (bfqd->tot_rq_in_driver >= limit)
4691 		return NULL;
4692 
4693 	/*
4694 	 * Linear search of the source queue for injection; but, with
4695 	 * a high probability, very few steps are needed to find a
4696 	 * candidate queue, i.e., a queue with enough budget left for
4697 	 * its next request. In fact:
4698 	 * - BFQ dynamically updates the budget of every queue so as
4699 	 *   to accommodate the expected backlog of the queue;
4700 	 * - if a queue gets all its requests dispatched as injected
4701 	 *   service, then the queue is removed from the active list
4702 	 *   (and re-added only if it gets new requests, but then it
4703 	 *   is assigned again enough budget for its new backlog).
4704 	 */
4705 	for (i = 0; i < bfqd->num_actuators; i++) {
4706 		list_for_each_entry(bfqq, &bfqd->active_list[i], bfqq_list)
4707 			if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4708 				(in_serv_always_inject || bfqq->wr_coeff > 1) &&
4709 				bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4710 				bfq_bfqq_budget_left(bfqq)) {
4711 			/*
4712 			 * Allow for only one large in-flight request
4713 			 * on non-rotational devices, for the
4714 			 * following reason. On non-rotationl drives,
4715 			 * large requests take much longer than
4716 			 * smaller requests to be served. In addition,
4717 			 * the drive prefers to serve large requests
4718 			 * w.r.t. to small ones, if it can choose. So,
4719 			 * having more than one large requests queued
4720 			 * in the drive may easily make the next first
4721 			 * request of the in-service queue wait for so
4722 			 * long to break bfqq's service guarantees. On
4723 			 * the bright side, large requests let the
4724 			 * drive reach a very high throughput, even if
4725 			 * there is only one in-flight large request
4726 			 * at a time.
4727 			 */
4728 			if (blk_queue_nonrot(bfqd->queue) &&
4729 			    blk_rq_sectors(bfqq->next_rq) >=
4730 			    BFQQ_SECT_THR_NONROT &&
4731 			    bfqd->tot_rq_in_driver >= 1)
4732 				continue;
4733 			else {
4734 				bfqd->rqs_injected = true;
4735 				return bfqq;
4736 			}
4737 		}
4738 	}
4739 
4740 	return NULL;
4741 }
4742 
4743 static struct bfq_queue *
4744 bfq_find_active_bfqq_for_actuator(struct bfq_data *bfqd, int idx)
4745 {
4746 	struct bfq_queue *bfqq;
4747 
4748 	if (bfqd->in_service_queue &&
4749 	    bfqd->in_service_queue->actuator_idx == idx)
4750 		return bfqd->in_service_queue;
4751 
4752 	list_for_each_entry(bfqq, &bfqd->active_list[idx], bfqq_list) {
4753 		if (!RB_EMPTY_ROOT(&bfqq->sort_list) &&
4754 			bfq_serv_to_charge(bfqq->next_rq, bfqq) <=
4755 				bfq_bfqq_budget_left(bfqq)) {
4756 			return bfqq;
4757 		}
4758 	}
4759 
4760 	return NULL;
4761 }
4762 
4763 /*
4764  * Perform a linear scan of each actuator, until an actuator is found
4765  * for which the following three conditions hold: the load of the
4766  * actuator is below the threshold (see comments on
4767  * actuator_load_threshold for details) and lower than that of the
4768  * next actuator (comments on this extra condition below), and there
4769  * is a queue that contains I/O for that actuator. On success, return
4770  * that queue.
4771  *
4772  * Performing a plain linear scan entails a prioritization among
4773  * actuators. The extra condition above breaks this prioritization and
4774  * tends to distribute injection uniformly across actuators.
4775  */
4776 static struct bfq_queue *
4777 bfq_find_bfqq_for_underused_actuator(struct bfq_data *bfqd)
4778 {
4779 	int i;
4780 
4781 	for (i = 0 ; i < bfqd->num_actuators; i++) {
4782 		if (bfqd->rq_in_driver[i] < bfqd->actuator_load_threshold &&
4783 		    (i == bfqd->num_actuators - 1 ||
4784 		     bfqd->rq_in_driver[i] < bfqd->rq_in_driver[i+1])) {
4785 			struct bfq_queue *bfqq =
4786 				bfq_find_active_bfqq_for_actuator(bfqd, i);
4787 
4788 			if (bfqq)
4789 				return bfqq;
4790 		}
4791 	}
4792 
4793 	return NULL;
4794 }
4795 
4796 
4797 /*
4798  * Select a queue for service.  If we have a current queue in service,
4799  * check whether to continue servicing it, or retrieve and set a new one.
4800  */
4801 static struct bfq_queue *bfq_select_queue(struct bfq_data *bfqd)
4802 {
4803 	struct bfq_queue *bfqq, *inject_bfqq;
4804 	struct request *next_rq;
4805 	enum bfqq_expiration reason = BFQQE_BUDGET_TIMEOUT;
4806 
4807 	bfqq = bfqd->in_service_queue;
4808 	if (!bfqq)
4809 		goto new_queue;
4810 
4811 	bfq_log_bfqq(bfqd, bfqq, "select_queue: already in-service queue");
4812 
4813 	/*
4814 	 * Do not expire bfqq for budget timeout if bfqq may be about
4815 	 * to enjoy device idling. The reason why, in this case, we
4816 	 * prevent bfqq from expiring is the same as in the comments
4817 	 * on the case where bfq_bfqq_must_idle() returns true, in
4818 	 * bfq_completed_request().
4819 	 */
4820 	if (bfq_may_expire_for_budg_timeout(bfqq) &&
4821 	    !bfq_bfqq_must_idle(bfqq))
4822 		goto expire;
4823 
4824 check_queue:
4825 	/*
4826 	 *  If some actuator is underutilized, but the in-service
4827 	 *  queue does not contain I/O for that actuator, then try to
4828 	 *  inject I/O for that actuator.
4829 	 */
4830 	inject_bfqq = bfq_find_bfqq_for_underused_actuator(bfqd);
4831 	if (inject_bfqq && inject_bfqq != bfqq)
4832 		return inject_bfqq;
4833 
4834 	/*
4835 	 * This loop is rarely executed more than once. Even when it
4836 	 * happens, it is much more convenient to re-execute this loop
4837 	 * than to return NULL and trigger a new dispatch to get a
4838 	 * request served.
4839 	 */
4840 	next_rq = bfqq->next_rq;
4841 	/*
4842 	 * If bfqq has requests queued and it has enough budget left to
4843 	 * serve them, keep the queue, otherwise expire it.
4844 	 */
4845 	if (next_rq) {
4846 		if (bfq_serv_to_charge(next_rq, bfqq) >
4847 			bfq_bfqq_budget_left(bfqq)) {
4848 			/*
4849 			 * Expire the queue for budget exhaustion,
4850 			 * which makes sure that the next budget is
4851 			 * enough to serve the next request, even if
4852 			 * it comes from the fifo expired path.
4853 			 */
4854 			reason = BFQQE_BUDGET_EXHAUSTED;
4855 			goto expire;
4856 		} else {
4857 			/*
4858 			 * The idle timer may be pending because we may
4859 			 * not disable disk idling even when a new request
4860 			 * arrives.
4861 			 */
4862 			if (bfq_bfqq_wait_request(bfqq)) {
4863 				/*
4864 				 * If we get here: 1) at least a new request
4865 				 * has arrived but we have not disabled the
4866 				 * timer because the request was too small,
4867 				 * 2) then the block layer has unplugged
4868 				 * the device, causing the dispatch to be
4869 				 * invoked.
4870 				 *
4871 				 * Since the device is unplugged, now the
4872 				 * requests are probably large enough to
4873 				 * provide a reasonable throughput.
4874 				 * So we disable idling.
4875 				 */
4876 				bfq_clear_bfqq_wait_request(bfqq);
4877 				hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
4878 			}
4879 			goto keep_queue;
4880 		}
4881 	}
4882 
4883 	/*
4884 	 * No requests pending. However, if the in-service queue is idling
4885 	 * for a new request, or has requests waiting for a completion and
4886 	 * may idle after their completion, then keep it anyway.
4887 	 *
4888 	 * Yet, inject service from other queues if it boosts
4889 	 * throughput and is possible.
4890 	 */
4891 	if (bfq_bfqq_wait_request(bfqq) ||
4892 	    (bfqq->dispatched != 0 && bfq_better_to_idle(bfqq))) {
4893 		unsigned int act_idx = bfqq->actuator_idx;
4894 		struct bfq_queue *async_bfqq = NULL;
4895 		struct bfq_queue *blocked_bfqq =
4896 			!hlist_empty(&bfqq->woken_list) ?
4897 			container_of(bfqq->woken_list.first,
4898 				     struct bfq_queue,
4899 				     woken_list_node)
4900 			: NULL;
4901 
4902 		if (bfqq->bic && bfqq->bic->bfqq[0][act_idx] &&
4903 		    bfq_bfqq_busy(bfqq->bic->bfqq[0][act_idx]) &&
4904 		    bfqq->bic->bfqq[0][act_idx]->next_rq)
4905 			async_bfqq = bfqq->bic->bfqq[0][act_idx];
4906 		/*
4907 		 * The next four mutually-exclusive ifs decide
4908 		 * whether to try injection, and choose the queue to
4909 		 * pick an I/O request from.
4910 		 *
4911 		 * The first if checks whether the process associated
4912 		 * with bfqq has also async I/O pending. If so, it
4913 		 * injects such I/O unconditionally. Injecting async
4914 		 * I/O from the same process can cause no harm to the
4915 		 * process. On the contrary, it can only increase
4916 		 * bandwidth and reduce latency for the process.
4917 		 *
4918 		 * The second if checks whether there happens to be a
4919 		 * non-empty waker queue for bfqq, i.e., a queue whose
4920 		 * I/O needs to be completed for bfqq to receive new
4921 		 * I/O. This happens, e.g., if bfqq is associated with
4922 		 * a process that does some sync. A sync generates
4923 		 * extra blocking I/O, which must be completed before
4924 		 * the process associated with bfqq can go on with its
4925 		 * I/O. If the I/O of the waker queue is not served,
4926 		 * then bfqq remains empty, and no I/O is dispatched,
4927 		 * until the idle timeout fires for bfqq. This is
4928 		 * likely to result in lower bandwidth and higher
4929 		 * latencies for bfqq, and in a severe loss of total
4930 		 * throughput. The best action to take is therefore to
4931 		 * serve the waker queue as soon as possible. So do it
4932 		 * (without relying on the third alternative below for
4933 		 * eventually serving waker_bfqq's I/O; see the last
4934 		 * paragraph for further details). This systematic
4935 		 * injection of I/O from the waker queue does not
4936 		 * cause any delay to bfqq's I/O. On the contrary,
4937 		 * next bfqq's I/O is brought forward dramatically,
4938 		 * for it is not blocked for milliseconds.
4939 		 *
4940 		 * The third if checks whether there is a queue woken
4941 		 * by bfqq, and currently with pending I/O. Such a
4942 		 * woken queue does not steal bandwidth from bfqq,
4943 		 * because it remains soon without I/O if bfqq is not
4944 		 * served. So there is virtually no risk of loss of
4945 		 * bandwidth for bfqq if this woken queue has I/O
4946 		 * dispatched while bfqq is waiting for new I/O.
4947 		 *
4948 		 * The fourth if checks whether bfqq is a queue for
4949 		 * which it is better to avoid injection. It is so if
4950 		 * bfqq delivers more throughput when served without
4951 		 * any further I/O from other queues in the middle, or
4952 		 * if the service times of bfqq's I/O requests both
4953 		 * count more than overall throughput, and may be
4954 		 * easily increased by injection (this happens if bfqq
4955 		 * has a short think time). If none of these
4956 		 * conditions holds, then a candidate queue for
4957 		 * injection is looked for through
4958 		 * bfq_choose_bfqq_for_injection(). Note that the
4959 		 * latter may return NULL (for example if the inject
4960 		 * limit for bfqq is currently 0).
4961 		 *
4962 		 * NOTE: motivation for the second alternative
4963 		 *
4964 		 * Thanks to the way the inject limit is updated in
4965 		 * bfq_update_has_short_ttime(), it is rather likely
4966 		 * that, if I/O is being plugged for bfqq and the
4967 		 * waker queue has pending I/O requests that are
4968 		 * blocking bfqq's I/O, then the fourth alternative
4969 		 * above lets the waker queue get served before the
4970 		 * I/O-plugging timeout fires. So one may deem the
4971 		 * second alternative superfluous. It is not, because
4972 		 * the fourth alternative may be way less effective in
4973 		 * case of a synchronization. For two main
4974 		 * reasons. First, throughput may be low because the
4975 		 * inject limit may be too low to guarantee the same
4976 		 * amount of injected I/O, from the waker queue or
4977 		 * other queues, that the second alternative
4978 		 * guarantees (the second alternative unconditionally
4979 		 * injects a pending I/O request of the waker queue
4980 		 * for each bfq_dispatch_request()). Second, with the
4981 		 * fourth alternative, the duration of the plugging,
4982 		 * i.e., the time before bfqq finally receives new I/O,
4983 		 * may not be minimized, because the waker queue may
4984 		 * happen to be served only after other queues.
4985 		 */
4986 		if (async_bfqq &&
4987 		    icq_to_bic(async_bfqq->next_rq->elv.icq) == bfqq->bic &&
4988 		    bfq_serv_to_charge(async_bfqq->next_rq, async_bfqq) <=
4989 		    bfq_bfqq_budget_left(async_bfqq))
4990 			bfqq = async_bfqq;
4991 		else if (bfqq->waker_bfqq &&
4992 			   bfq_bfqq_busy(bfqq->waker_bfqq) &&
4993 			   bfqq->waker_bfqq->next_rq &&
4994 			   bfq_serv_to_charge(bfqq->waker_bfqq->next_rq,
4995 					      bfqq->waker_bfqq) <=
4996 			   bfq_bfqq_budget_left(bfqq->waker_bfqq)
4997 			)
4998 			bfqq = bfqq->waker_bfqq;
4999 		else if (blocked_bfqq &&
5000 			   bfq_bfqq_busy(blocked_bfqq) &&
5001 			   blocked_bfqq->next_rq &&
5002 			   bfq_serv_to_charge(blocked_bfqq->next_rq,
5003 					      blocked_bfqq) <=
5004 			   bfq_bfqq_budget_left(blocked_bfqq)
5005 			)
5006 			bfqq = blocked_bfqq;
5007 		else if (!idling_boosts_thr_without_issues(bfqd, bfqq) &&
5008 			 (bfqq->wr_coeff == 1 || bfqd->wr_busy_queues > 1 ||
5009 			  !bfq_bfqq_has_short_ttime(bfqq)))
5010 			bfqq = bfq_choose_bfqq_for_injection(bfqd);
5011 		else
5012 			bfqq = NULL;
5013 
5014 		goto keep_queue;
5015 	}
5016 
5017 	reason = BFQQE_NO_MORE_REQUESTS;
5018 expire:
5019 	bfq_bfqq_expire(bfqd, bfqq, false, reason);
5020 new_queue:
5021 	bfqq = bfq_set_in_service_queue(bfqd);
5022 	if (bfqq) {
5023 		bfq_log_bfqq(bfqd, bfqq, "select_queue: checking new queue");
5024 		goto check_queue;
5025 	}
5026 keep_queue:
5027 	if (bfqq)
5028 		bfq_log_bfqq(bfqd, bfqq, "select_queue: returned this queue");
5029 	else
5030 		bfq_log(bfqd, "select_queue: no queue returned");
5031 
5032 	return bfqq;
5033 }
5034 
5035 static void bfq_update_wr_data(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5036 {
5037 	struct bfq_entity *entity = &bfqq->entity;
5038 
5039 	if (bfqq->wr_coeff > 1) { /* queue is being weight-raised */
5040 		bfq_log_bfqq(bfqd, bfqq,
5041 			"raising period dur %u/%u msec, old coeff %u, w %d(%d)",
5042 			jiffies_to_msecs(jiffies - bfqq->last_wr_start_finish),
5043 			jiffies_to_msecs(bfqq->wr_cur_max_time),
5044 			bfqq->wr_coeff,
5045 			bfqq->entity.weight, bfqq->entity.orig_weight);
5046 
5047 		if (entity->prio_changed)
5048 			bfq_log_bfqq(bfqd, bfqq, "WARN: pending prio change");
5049 
5050 		/*
5051 		 * If the queue was activated in a burst, or too much
5052 		 * time has elapsed from the beginning of this
5053 		 * weight-raising period, then end weight raising.
5054 		 */
5055 		if (bfq_bfqq_in_large_burst(bfqq))
5056 			bfq_bfqq_end_wr(bfqq);
5057 		else if (time_is_before_jiffies(bfqq->last_wr_start_finish +
5058 						bfqq->wr_cur_max_time)) {
5059 			if (bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time ||
5060 			time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5061 					       bfq_wr_duration(bfqd))) {
5062 				/*
5063 				 * Either in interactive weight
5064 				 * raising, or in soft_rt weight
5065 				 * raising with the
5066 				 * interactive-weight-raising period
5067 				 * elapsed (so no switch back to
5068 				 * interactive weight raising).
5069 				 */
5070 				bfq_bfqq_end_wr(bfqq);
5071 			} else { /*
5072 				  * soft_rt finishing while still in
5073 				  * interactive period, switch back to
5074 				  * interactive weight raising
5075 				  */
5076 				switch_back_to_interactive_wr(bfqq, bfqd);
5077 				bfqq->entity.prio_changed = 1;
5078 			}
5079 		}
5080 		if (bfqq->wr_coeff > 1 &&
5081 		    bfqq->wr_cur_max_time != bfqd->bfq_wr_rt_max_time &&
5082 		    bfqq->service_from_wr > max_service_from_wr) {
5083 			/* see comments on max_service_from_wr */
5084 			bfq_bfqq_end_wr(bfqq);
5085 		}
5086 	}
5087 	/*
5088 	 * To improve latency (for this or other queues), immediately
5089 	 * update weight both if it must be raised and if it must be
5090 	 * lowered. Since, entity may be on some active tree here, and
5091 	 * might have a pending change of its ioprio class, invoke
5092 	 * next function with the last parameter unset (see the
5093 	 * comments on the function).
5094 	 */
5095 	if ((entity->weight > entity->orig_weight) != (bfqq->wr_coeff > 1))
5096 		__bfq_entity_update_weight_prio(bfq_entity_service_tree(entity),
5097 						entity, false);
5098 }
5099 
5100 /*
5101  * Dispatch next request from bfqq.
5102  */
5103 static struct request *bfq_dispatch_rq_from_bfqq(struct bfq_data *bfqd,
5104 						 struct bfq_queue *bfqq)
5105 {
5106 	struct request *rq = bfqq->next_rq;
5107 	unsigned long service_to_charge;
5108 
5109 	service_to_charge = bfq_serv_to_charge(rq, bfqq);
5110 
5111 	bfq_bfqq_served(bfqq, service_to_charge);
5112 
5113 	if (bfqq == bfqd->in_service_queue && bfqd->wait_dispatch) {
5114 		bfqd->wait_dispatch = false;
5115 		bfqd->waited_rq = rq;
5116 	}
5117 
5118 	bfq_dispatch_remove(bfqd->queue, rq);
5119 
5120 	if (bfqq != bfqd->in_service_queue)
5121 		return rq;
5122 
5123 	/*
5124 	 * If weight raising has to terminate for bfqq, then next
5125 	 * function causes an immediate update of bfqq's weight,
5126 	 * without waiting for next activation. As a consequence, on
5127 	 * expiration, bfqq will be timestamped as if has never been
5128 	 * weight-raised during this service slot, even if it has
5129 	 * received part or even most of the service as a
5130 	 * weight-raised queue. This inflates bfqq's timestamps, which
5131 	 * is beneficial, as bfqq is then more willing to leave the
5132 	 * device immediately to possible other weight-raised queues.
5133 	 */
5134 	bfq_update_wr_data(bfqd, bfqq);
5135 
5136 	/*
5137 	 * Expire bfqq, pretending that its budget expired, if bfqq
5138 	 * belongs to CLASS_IDLE and other queues are waiting for
5139 	 * service.
5140 	 */
5141 	if (bfq_tot_busy_queues(bfqd) > 1 && bfq_class_idle(bfqq))
5142 		bfq_bfqq_expire(bfqd, bfqq, false, BFQQE_BUDGET_EXHAUSTED);
5143 
5144 	return rq;
5145 }
5146 
5147 static bool bfq_has_work(struct blk_mq_hw_ctx *hctx)
5148 {
5149 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5150 
5151 	/*
5152 	 * Avoiding lock: a race on bfqd->queued should cause at
5153 	 * most a call to dispatch for nothing
5154 	 */
5155 	return !list_empty_careful(&bfqd->dispatch) ||
5156 		READ_ONCE(bfqd->queued);
5157 }
5158 
5159 static struct request *__bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5160 {
5161 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5162 	struct request *rq = NULL;
5163 	struct bfq_queue *bfqq = NULL;
5164 
5165 	if (!list_empty(&bfqd->dispatch)) {
5166 		rq = list_first_entry(&bfqd->dispatch, struct request,
5167 				      queuelist);
5168 		list_del_init(&rq->queuelist);
5169 
5170 		bfqq = RQ_BFQQ(rq);
5171 
5172 		if (bfqq) {
5173 			/*
5174 			 * Increment counters here, because this
5175 			 * dispatch does not follow the standard
5176 			 * dispatch flow (where counters are
5177 			 * incremented)
5178 			 */
5179 			bfqq->dispatched++;
5180 
5181 			goto inc_in_driver_start_rq;
5182 		}
5183 
5184 		/*
5185 		 * We exploit the bfq_finish_requeue_request hook to
5186 		 * decrement tot_rq_in_driver, but
5187 		 * bfq_finish_requeue_request will not be invoked on
5188 		 * this request. So, to avoid unbalance, just start
5189 		 * this request, without incrementing tot_rq_in_driver. As
5190 		 * a negative consequence, tot_rq_in_driver is deceptively
5191 		 * lower than it should be while this request is in
5192 		 * service. This may cause bfq_schedule_dispatch to be
5193 		 * invoked uselessly.
5194 		 *
5195 		 * As for implementing an exact solution, the
5196 		 * bfq_finish_requeue_request hook, if defined, is
5197 		 * probably invoked also on this request. So, by
5198 		 * exploiting this hook, we could 1) increment
5199 		 * tot_rq_in_driver here, and 2) decrement it in
5200 		 * bfq_finish_requeue_request. Such a solution would
5201 		 * let the value of the counter be always accurate,
5202 		 * but it would entail using an extra interface
5203 		 * function. This cost seems higher than the benefit,
5204 		 * being the frequency of non-elevator-private
5205 		 * requests very low.
5206 		 */
5207 		goto start_rq;
5208 	}
5209 
5210 	bfq_log(bfqd, "dispatch requests: %d busy queues",
5211 		bfq_tot_busy_queues(bfqd));
5212 
5213 	if (bfq_tot_busy_queues(bfqd) == 0)
5214 		goto exit;
5215 
5216 	/*
5217 	 * Force device to serve one request at a time if
5218 	 * strict_guarantees is true. Forcing this service scheme is
5219 	 * currently the ONLY way to guarantee that the request
5220 	 * service order enforced by the scheduler is respected by a
5221 	 * queueing device. Otherwise the device is free even to make
5222 	 * some unlucky request wait for as long as the device
5223 	 * wishes.
5224 	 *
5225 	 * Of course, serving one request at a time may cause loss of
5226 	 * throughput.
5227 	 */
5228 	if (bfqd->strict_guarantees && bfqd->tot_rq_in_driver > 0)
5229 		goto exit;
5230 
5231 	bfqq = bfq_select_queue(bfqd);
5232 	if (!bfqq)
5233 		goto exit;
5234 
5235 	rq = bfq_dispatch_rq_from_bfqq(bfqd, bfqq);
5236 
5237 	if (rq) {
5238 inc_in_driver_start_rq:
5239 		bfqd->rq_in_driver[bfqq->actuator_idx]++;
5240 		bfqd->tot_rq_in_driver++;
5241 start_rq:
5242 		rq->rq_flags |= RQF_STARTED;
5243 	}
5244 exit:
5245 	return rq;
5246 }
5247 
5248 #ifdef CONFIG_BFQ_CGROUP_DEBUG
5249 static void bfq_update_dispatch_stats(struct request_queue *q,
5250 				      struct request *rq,
5251 				      struct bfq_queue *in_serv_queue,
5252 				      bool idle_timer_disabled)
5253 {
5254 	struct bfq_queue *bfqq = rq ? RQ_BFQQ(rq) : NULL;
5255 
5256 	if (!idle_timer_disabled && !bfqq)
5257 		return;
5258 
5259 	/*
5260 	 * rq and bfqq are guaranteed to exist until this function
5261 	 * ends, for the following reasons. First, rq can be
5262 	 * dispatched to the device, and then can be completed and
5263 	 * freed, only after this function ends. Second, rq cannot be
5264 	 * merged (and thus freed because of a merge) any longer,
5265 	 * because it has already started. Thus rq cannot be freed
5266 	 * before this function ends, and, since rq has a reference to
5267 	 * bfqq, the same guarantee holds for bfqq too.
5268 	 *
5269 	 * In addition, the following queue lock guarantees that
5270 	 * bfqq_group(bfqq) exists as well.
5271 	 */
5272 	spin_lock_irq(&q->queue_lock);
5273 	if (idle_timer_disabled)
5274 		/*
5275 		 * Since the idle timer has been disabled,
5276 		 * in_serv_queue contained some request when
5277 		 * __bfq_dispatch_request was invoked above, which
5278 		 * implies that rq was picked exactly from
5279 		 * in_serv_queue. Thus in_serv_queue == bfqq, and is
5280 		 * therefore guaranteed to exist because of the above
5281 		 * arguments.
5282 		 */
5283 		bfqg_stats_update_idle_time(bfqq_group(in_serv_queue));
5284 	if (bfqq) {
5285 		struct bfq_group *bfqg = bfqq_group(bfqq);
5286 
5287 		bfqg_stats_update_avg_queue_size(bfqg);
5288 		bfqg_stats_set_start_empty_time(bfqg);
5289 		bfqg_stats_update_io_remove(bfqg, rq->cmd_flags);
5290 	}
5291 	spin_unlock_irq(&q->queue_lock);
5292 }
5293 #else
5294 static inline void bfq_update_dispatch_stats(struct request_queue *q,
5295 					     struct request *rq,
5296 					     struct bfq_queue *in_serv_queue,
5297 					     bool idle_timer_disabled) {}
5298 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
5299 
5300 static struct request *bfq_dispatch_request(struct blk_mq_hw_ctx *hctx)
5301 {
5302 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
5303 	struct request *rq;
5304 	struct bfq_queue *in_serv_queue;
5305 	bool waiting_rq, idle_timer_disabled = false;
5306 
5307 	spin_lock_irq(&bfqd->lock);
5308 
5309 	in_serv_queue = bfqd->in_service_queue;
5310 	waiting_rq = in_serv_queue && bfq_bfqq_wait_request(in_serv_queue);
5311 
5312 	rq = __bfq_dispatch_request(hctx);
5313 	if (in_serv_queue == bfqd->in_service_queue) {
5314 		idle_timer_disabled =
5315 			waiting_rq && !bfq_bfqq_wait_request(in_serv_queue);
5316 	}
5317 
5318 	spin_unlock_irq(&bfqd->lock);
5319 	bfq_update_dispatch_stats(hctx->queue, rq,
5320 			idle_timer_disabled ? in_serv_queue : NULL,
5321 				idle_timer_disabled);
5322 
5323 	return rq;
5324 }
5325 
5326 /*
5327  * Task holds one reference to the queue, dropped when task exits.  Each rq
5328  * in-flight on this queue also holds a reference, dropped when rq is freed.
5329  *
5330  * Scheduler lock must be held here. Recall not to use bfqq after calling
5331  * this function on it.
5332  */
5333 void bfq_put_queue(struct bfq_queue *bfqq)
5334 {
5335 	struct bfq_queue *item;
5336 	struct hlist_node *n;
5337 	struct bfq_group *bfqg = bfqq_group(bfqq);
5338 
5339 	bfq_log_bfqq(bfqq->bfqd, bfqq, "put_queue: %p %d", bfqq, bfqq->ref);
5340 
5341 	bfqq->ref--;
5342 	if (bfqq->ref)
5343 		return;
5344 
5345 	if (!hlist_unhashed(&bfqq->burst_list_node)) {
5346 		hlist_del_init(&bfqq->burst_list_node);
5347 		/*
5348 		 * Decrement also burst size after the removal, if the
5349 		 * process associated with bfqq is exiting, and thus
5350 		 * does not contribute to the burst any longer. This
5351 		 * decrement helps filter out false positives of large
5352 		 * bursts, when some short-lived process (often due to
5353 		 * the execution of commands by some service) happens
5354 		 * to start and exit while a complex application is
5355 		 * starting, and thus spawning several processes that
5356 		 * do I/O (and that *must not* be treated as a large
5357 		 * burst, see comments on bfq_handle_burst).
5358 		 *
5359 		 * In particular, the decrement is performed only if:
5360 		 * 1) bfqq is not a merged queue, because, if it is,
5361 		 * then this free of bfqq is not triggered by the exit
5362 		 * of the process bfqq is associated with, but exactly
5363 		 * by the fact that bfqq has just been merged.
5364 		 * 2) burst_size is greater than 0, to handle
5365 		 * unbalanced decrements. Unbalanced decrements may
5366 		 * happen in te following case: bfqq is inserted into
5367 		 * the current burst list--without incrementing
5368 		 * bust_size--because of a split, but the current
5369 		 * burst list is not the burst list bfqq belonged to
5370 		 * (see comments on the case of a split in
5371 		 * bfq_set_request).
5372 		 */
5373 		if (bfqq->bic && bfqq->bfqd->burst_size > 0)
5374 			bfqq->bfqd->burst_size--;
5375 	}
5376 
5377 	/*
5378 	 * bfqq does not exist any longer, so it cannot be woken by
5379 	 * any other queue, and cannot wake any other queue. Then bfqq
5380 	 * must be removed from the woken list of its possible waker
5381 	 * queue, and all queues in the woken list of bfqq must stop
5382 	 * having a waker queue. Strictly speaking, these updates
5383 	 * should be performed when bfqq remains with no I/O source
5384 	 * attached to it, which happens before bfqq gets freed. In
5385 	 * particular, this happens when the last process associated
5386 	 * with bfqq exits or gets associated with a different
5387 	 * queue. However, both events lead to bfqq being freed soon,
5388 	 * and dangling references would come out only after bfqq gets
5389 	 * freed. So these updates are done here, as a simple and safe
5390 	 * way to handle all cases.
5391 	 */
5392 	/* remove bfqq from woken list */
5393 	if (!hlist_unhashed(&bfqq->woken_list_node))
5394 		hlist_del_init(&bfqq->woken_list_node);
5395 
5396 	/* reset waker for all queues in woken list */
5397 	hlist_for_each_entry_safe(item, n, &bfqq->woken_list,
5398 				  woken_list_node) {
5399 		item->waker_bfqq = NULL;
5400 		hlist_del_init(&item->woken_list_node);
5401 	}
5402 
5403 	if (bfqq->bfqd->last_completed_rq_bfqq == bfqq)
5404 		bfqq->bfqd->last_completed_rq_bfqq = NULL;
5405 
5406 	kmem_cache_free(bfq_pool, bfqq);
5407 	bfqg_and_blkg_put(bfqg);
5408 }
5409 
5410 static void bfq_put_stable_ref(struct bfq_queue *bfqq)
5411 {
5412 	bfqq->stable_ref--;
5413 	bfq_put_queue(bfqq);
5414 }
5415 
5416 void bfq_put_cooperator(struct bfq_queue *bfqq)
5417 {
5418 	struct bfq_queue *__bfqq, *next;
5419 
5420 	/*
5421 	 * If this queue was scheduled to merge with another queue, be
5422 	 * sure to drop the reference taken on that queue (and others in
5423 	 * the merge chain). See bfq_setup_merge and bfq_merge_bfqqs.
5424 	 */
5425 	__bfqq = bfqq->new_bfqq;
5426 	while (__bfqq) {
5427 		next = __bfqq->new_bfqq;
5428 		bfq_put_queue(__bfqq);
5429 		__bfqq = next;
5430 	}
5431 }
5432 
5433 static void bfq_exit_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq)
5434 {
5435 	if (bfqq == bfqd->in_service_queue) {
5436 		__bfq_bfqq_expire(bfqd, bfqq, BFQQE_BUDGET_TIMEOUT);
5437 		bfq_schedule_dispatch(bfqd);
5438 	}
5439 
5440 	bfq_log_bfqq(bfqd, bfqq, "exit_bfqq: %p, %d", bfqq, bfqq->ref);
5441 
5442 	bfq_put_cooperator(bfqq);
5443 
5444 	bfq_release_process_ref(bfqd, bfqq);
5445 }
5446 
5447 static void bfq_exit_icq_bfqq(struct bfq_io_cq *bic, bool is_sync,
5448 			      unsigned int actuator_idx)
5449 {
5450 	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync, actuator_idx);
5451 	struct bfq_data *bfqd;
5452 
5453 	if (bfqq)
5454 		bfqd = bfqq->bfqd; /* NULL if scheduler already exited */
5455 
5456 	if (bfqq && bfqd) {
5457 		bic_set_bfqq(bic, NULL, is_sync, actuator_idx);
5458 		bfq_exit_bfqq(bfqd, bfqq);
5459 	}
5460 }
5461 
5462 static void bfq_exit_icq(struct io_cq *icq)
5463 {
5464 	struct bfq_io_cq *bic = icq_to_bic(icq);
5465 	struct bfq_data *bfqd = bic_to_bfqd(bic);
5466 	unsigned long flags;
5467 	unsigned int act_idx;
5468 	/*
5469 	 * If bfqd and thus bfqd->num_actuators is not available any
5470 	 * longer, then cycle over all possible per-actuator bfqqs in
5471 	 * next loop. We rely on bic being zeroed on creation, and
5472 	 * therefore on its unused per-actuator fields being NULL.
5473 	 */
5474 	unsigned int num_actuators = BFQ_MAX_ACTUATORS;
5475 	struct bfq_iocq_bfqq_data *bfqq_data = bic->bfqq_data;
5476 
5477 	/*
5478 	 * bfqd is NULL if scheduler already exited, and in that case
5479 	 * this is the last time these queues are accessed.
5480 	 */
5481 	if (bfqd) {
5482 		spin_lock_irqsave(&bfqd->lock, flags);
5483 		num_actuators = bfqd->num_actuators;
5484 	}
5485 
5486 	for (act_idx = 0; act_idx < num_actuators; act_idx++) {
5487 		if (bfqq_data[act_idx].stable_merge_bfqq)
5488 			bfq_put_stable_ref(bfqq_data[act_idx].stable_merge_bfqq);
5489 
5490 		bfq_exit_icq_bfqq(bic, true, act_idx);
5491 		bfq_exit_icq_bfqq(bic, false, act_idx);
5492 	}
5493 
5494 	if (bfqd)
5495 		spin_unlock_irqrestore(&bfqd->lock, flags);
5496 }
5497 
5498 /*
5499  * Update the entity prio values; note that the new values will not
5500  * be used until the next (re)activation.
5501  */
5502 static void
5503 bfq_set_next_ioprio_data(struct bfq_queue *bfqq, struct bfq_io_cq *bic)
5504 {
5505 	struct task_struct *tsk = current;
5506 	int ioprio_class;
5507 	struct bfq_data *bfqd = bfqq->bfqd;
5508 
5509 	if (!bfqd)
5510 		return;
5511 
5512 	ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5513 	switch (ioprio_class) {
5514 	default:
5515 		pr_err("bdi %s: bfq: bad prio class %d\n",
5516 			bdi_dev_name(bfqq->bfqd->queue->disk->bdi),
5517 			ioprio_class);
5518 		fallthrough;
5519 	case IOPRIO_CLASS_NONE:
5520 		/*
5521 		 * No prio set, inherit CPU scheduling settings.
5522 		 */
5523 		bfqq->new_ioprio = task_nice_ioprio(tsk);
5524 		bfqq->new_ioprio_class = task_nice_ioclass(tsk);
5525 		break;
5526 	case IOPRIO_CLASS_RT:
5527 		bfqq->new_ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5528 		bfqq->new_ioprio_class = IOPRIO_CLASS_RT;
5529 		break;
5530 	case IOPRIO_CLASS_BE:
5531 		bfqq->new_ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5532 		bfqq->new_ioprio_class = IOPRIO_CLASS_BE;
5533 		break;
5534 	case IOPRIO_CLASS_IDLE:
5535 		bfqq->new_ioprio_class = IOPRIO_CLASS_IDLE;
5536 		bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5537 		break;
5538 	}
5539 
5540 	if (bfqq->new_ioprio >= IOPRIO_NR_LEVELS) {
5541 		pr_crit("bfq_set_next_ioprio_data: new_ioprio %d\n",
5542 			bfqq->new_ioprio);
5543 		bfqq->new_ioprio = IOPRIO_NR_LEVELS - 1;
5544 	}
5545 
5546 	bfqq->entity.new_weight = bfq_ioprio_to_weight(bfqq->new_ioprio);
5547 	bfq_log_bfqq(bfqd, bfqq, "new_ioprio %d new_weight %d",
5548 		     bfqq->new_ioprio, bfqq->entity.new_weight);
5549 	bfqq->entity.prio_changed = 1;
5550 }
5551 
5552 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5553 				       struct bio *bio, bool is_sync,
5554 				       struct bfq_io_cq *bic,
5555 				       bool respawn);
5556 
5557 static void bfq_check_ioprio_change(struct bfq_io_cq *bic, struct bio *bio)
5558 {
5559 	struct bfq_data *bfqd = bic_to_bfqd(bic);
5560 	struct bfq_queue *bfqq;
5561 	int ioprio = bic->icq.ioc->ioprio;
5562 
5563 	/*
5564 	 * This condition may trigger on a newly created bic, be sure to
5565 	 * drop the lock before returning.
5566 	 */
5567 	if (unlikely(!bfqd) || likely(bic->ioprio == ioprio))
5568 		return;
5569 
5570 	bic->ioprio = ioprio;
5571 
5572 	bfqq = bic_to_bfqq(bic, false, bfq_actuator_index(bfqd, bio));
5573 	if (bfqq) {
5574 		struct bfq_queue *old_bfqq = bfqq;
5575 
5576 		bfqq = bfq_get_queue(bfqd, bio, false, bic, true);
5577 		bic_set_bfqq(bic, bfqq, false, bfq_actuator_index(bfqd, bio));
5578 		bfq_release_process_ref(bfqd, old_bfqq);
5579 	}
5580 
5581 	bfqq = bic_to_bfqq(bic, true, bfq_actuator_index(bfqd, bio));
5582 	if (bfqq)
5583 		bfq_set_next_ioprio_data(bfqq, bic);
5584 }
5585 
5586 static void bfq_init_bfqq(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5587 			  struct bfq_io_cq *bic, pid_t pid, int is_sync,
5588 			  unsigned int act_idx)
5589 {
5590 	u64 now_ns = ktime_get_ns();
5591 
5592 	bfqq->actuator_idx = act_idx;
5593 	RB_CLEAR_NODE(&bfqq->entity.rb_node);
5594 	INIT_LIST_HEAD(&bfqq->fifo);
5595 	INIT_HLIST_NODE(&bfqq->burst_list_node);
5596 	INIT_HLIST_NODE(&bfqq->woken_list_node);
5597 	INIT_HLIST_HEAD(&bfqq->woken_list);
5598 
5599 	bfqq->ref = 0;
5600 	bfqq->bfqd = bfqd;
5601 
5602 	if (bic)
5603 		bfq_set_next_ioprio_data(bfqq, bic);
5604 
5605 	if (is_sync) {
5606 		/*
5607 		 * No need to mark as has_short_ttime if in
5608 		 * idle_class, because no device idling is performed
5609 		 * for queues in idle class
5610 		 */
5611 		if (!bfq_class_idle(bfqq))
5612 			/* tentatively mark as has_short_ttime */
5613 			bfq_mark_bfqq_has_short_ttime(bfqq);
5614 		bfq_mark_bfqq_sync(bfqq);
5615 		bfq_mark_bfqq_just_created(bfqq);
5616 	} else
5617 		bfq_clear_bfqq_sync(bfqq);
5618 
5619 	/* set end request to minus infinity from now */
5620 	bfqq->ttime.last_end_request = now_ns + 1;
5621 
5622 	bfqq->creation_time = jiffies;
5623 
5624 	bfqq->io_start_time = now_ns;
5625 
5626 	bfq_mark_bfqq_IO_bound(bfqq);
5627 
5628 	bfqq->pid = pid;
5629 
5630 	/* Tentative initial value to trade off between thr and lat */
5631 	bfqq->max_budget = (2 * bfq_max_budget(bfqd)) / 3;
5632 	bfqq->budget_timeout = bfq_smallest_from_now();
5633 
5634 	bfqq->wr_coeff = 1;
5635 	bfqq->last_wr_start_finish = jiffies;
5636 	bfqq->wr_start_at_switch_to_srt = bfq_smallest_from_now();
5637 	bfqq->split_time = bfq_smallest_from_now();
5638 
5639 	/*
5640 	 * To not forget the possibly high bandwidth consumed by a
5641 	 * process/queue in the recent past,
5642 	 * bfq_bfqq_softrt_next_start() returns a value at least equal
5643 	 * to the current value of bfqq->soft_rt_next_start (see
5644 	 * comments on bfq_bfqq_softrt_next_start).  Set
5645 	 * soft_rt_next_start to now, to mean that bfqq has consumed
5646 	 * no bandwidth so far.
5647 	 */
5648 	bfqq->soft_rt_next_start = jiffies;
5649 
5650 	/* first request is almost certainly seeky */
5651 	bfqq->seek_history = 1;
5652 
5653 	bfqq->decrease_time_jif = jiffies;
5654 }
5655 
5656 static struct bfq_queue **bfq_async_queue_prio(struct bfq_data *bfqd,
5657 					       struct bfq_group *bfqg,
5658 					       int ioprio_class, int ioprio, int act_idx)
5659 {
5660 	switch (ioprio_class) {
5661 	case IOPRIO_CLASS_RT:
5662 		return &bfqg->async_bfqq[0][ioprio][act_idx];
5663 	case IOPRIO_CLASS_NONE:
5664 		ioprio = IOPRIO_BE_NORM;
5665 		fallthrough;
5666 	case IOPRIO_CLASS_BE:
5667 		return &bfqg->async_bfqq[1][ioprio][act_idx];
5668 	case IOPRIO_CLASS_IDLE:
5669 		return &bfqg->async_idle_bfqq[act_idx];
5670 	default:
5671 		return NULL;
5672 	}
5673 }
5674 
5675 static struct bfq_queue *
5676 bfq_do_early_stable_merge(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5677 			  struct bfq_io_cq *bic,
5678 			  struct bfq_queue *last_bfqq_created)
5679 {
5680 	unsigned int a_idx = last_bfqq_created->actuator_idx;
5681 	struct bfq_queue *new_bfqq =
5682 		bfq_setup_merge(bfqq, last_bfqq_created);
5683 
5684 	if (!new_bfqq)
5685 		return bfqq;
5686 
5687 	if (new_bfqq->bic)
5688 		new_bfqq->bic->bfqq_data[a_idx].stably_merged = true;
5689 	bic->bfqq_data[a_idx].stably_merged = true;
5690 
5691 	/*
5692 	 * Reusing merge functions. This implies that
5693 	 * bfqq->bic must be set too, for
5694 	 * bfq_merge_bfqqs to correctly save bfqq's
5695 	 * state before killing it.
5696 	 */
5697 	bfqq->bic = bic;
5698 	bfq_merge_bfqqs(bfqd, bic, bfqq, new_bfqq);
5699 
5700 	return new_bfqq;
5701 }
5702 
5703 /*
5704  * Many throughput-sensitive workloads are made of several parallel
5705  * I/O flows, with all flows generated by the same application, or
5706  * more generically by the same task (e.g., system boot). The most
5707  * counterproductive action with these workloads is plugging I/O
5708  * dispatch when one of the bfq_queues associated with these flows
5709  * remains temporarily empty.
5710  *
5711  * To avoid this plugging, BFQ has been using a burst-handling
5712  * mechanism for years now. This mechanism has proven effective for
5713  * throughput, and not detrimental for service guarantees. The
5714  * following function pushes this mechanism a little bit further,
5715  * basing on the following two facts.
5716  *
5717  * First, all the I/O flows of a the same application or task
5718  * contribute to the execution/completion of that common application
5719  * or task. So the performance figures that matter are total
5720  * throughput of the flows and task-wide I/O latency.  In particular,
5721  * these flows do not need to be protected from each other, in terms
5722  * of individual bandwidth or latency.
5723  *
5724  * Second, the above fact holds regardless of the number of flows.
5725  *
5726  * Putting these two facts together, this commits merges stably the
5727  * bfq_queues associated with these I/O flows, i.e., with the
5728  * processes that generate these IO/ flows, regardless of how many the
5729  * involved processes are.
5730  *
5731  * To decide whether a set of bfq_queues is actually associated with
5732  * the I/O flows of a common application or task, and to merge these
5733  * queues stably, this function operates as follows: given a bfq_queue,
5734  * say Q2, currently being created, and the last bfq_queue, say Q1,
5735  * created before Q2, Q2 is merged stably with Q1 if
5736  * - very little time has elapsed since when Q1 was created
5737  * - Q2 has the same ioprio as Q1
5738  * - Q2 belongs to the same group as Q1
5739  *
5740  * Merging bfq_queues also reduces scheduling overhead. A fio test
5741  * with ten random readers on /dev/nullb shows a throughput boost of
5742  * 40%, with a quadcore. Since BFQ's execution time amounts to ~50% of
5743  * the total per-request processing time, the above throughput boost
5744  * implies that BFQ's overhead is reduced by more than 50%.
5745  *
5746  * This new mechanism most certainly obsoletes the current
5747  * burst-handling heuristics. We keep those heuristics for the moment.
5748  */
5749 static struct bfq_queue *bfq_do_or_sched_stable_merge(struct bfq_data *bfqd,
5750 						      struct bfq_queue *bfqq,
5751 						      struct bfq_io_cq *bic)
5752 {
5753 	struct bfq_queue **source_bfqq = bfqq->entity.parent ?
5754 		&bfqq->entity.parent->last_bfqq_created :
5755 		&bfqd->last_bfqq_created;
5756 
5757 	struct bfq_queue *last_bfqq_created = *source_bfqq;
5758 
5759 	/*
5760 	 * If last_bfqq_created has not been set yet, then init it. If
5761 	 * it has been set already, but too long ago, then move it
5762 	 * forward to bfqq. Finally, move also if bfqq belongs to a
5763 	 * different group than last_bfqq_created, or if bfqq has a
5764 	 * different ioprio, ioprio_class or actuator_idx. If none of
5765 	 * these conditions holds true, then try an early stable merge
5766 	 * or schedule a delayed stable merge. As for the condition on
5767 	 * actuator_idx, the reason is that, if queues associated with
5768 	 * different actuators are merged, then control is lost on
5769 	 * each actuator. Therefore some actuator may be
5770 	 * underutilized, and throughput may decrease.
5771 	 *
5772 	 * A delayed merge is scheduled (instead of performing an
5773 	 * early merge), in case bfqq might soon prove to be more
5774 	 * throughput-beneficial if not merged. Currently this is
5775 	 * possible only if bfqd is rotational with no queueing. For
5776 	 * such a drive, not merging bfqq is better for throughput if
5777 	 * bfqq happens to contain sequential I/O. So, we wait a
5778 	 * little bit for enough I/O to flow through bfqq. After that,
5779 	 * if such an I/O is sequential, then the merge is
5780 	 * canceled. Otherwise the merge is finally performed.
5781 	 */
5782 	if (!last_bfqq_created ||
5783 	    time_before(last_bfqq_created->creation_time +
5784 			msecs_to_jiffies(bfq_activation_stable_merging),
5785 			bfqq->creation_time) ||
5786 		bfqq->entity.parent != last_bfqq_created->entity.parent ||
5787 		bfqq->ioprio != last_bfqq_created->ioprio ||
5788 		bfqq->ioprio_class != last_bfqq_created->ioprio_class ||
5789 		bfqq->actuator_idx != last_bfqq_created->actuator_idx)
5790 		*source_bfqq = bfqq;
5791 	else if (time_after_eq(last_bfqq_created->creation_time +
5792 				 bfqd->bfq_burst_interval,
5793 				 bfqq->creation_time)) {
5794 		if (likely(bfqd->nonrot_with_queueing))
5795 			/*
5796 			 * With this type of drive, leaving
5797 			 * bfqq alone may provide no
5798 			 * throughput benefits compared with
5799 			 * merging bfqq. So merge bfqq now.
5800 			 */
5801 			bfqq = bfq_do_early_stable_merge(bfqd, bfqq,
5802 							 bic,
5803 							 last_bfqq_created);
5804 		else { /* schedule tentative stable merge */
5805 			/*
5806 			 * get reference on last_bfqq_created,
5807 			 * to prevent it from being freed,
5808 			 * until we decide whether to merge
5809 			 */
5810 			last_bfqq_created->ref++;
5811 			/*
5812 			 * need to keep track of stable refs, to
5813 			 * compute process refs correctly
5814 			 */
5815 			last_bfqq_created->stable_ref++;
5816 			/*
5817 			 * Record the bfqq to merge to.
5818 			 */
5819 			bic->bfqq_data[last_bfqq_created->actuator_idx].stable_merge_bfqq =
5820 				last_bfqq_created;
5821 		}
5822 	}
5823 
5824 	return bfqq;
5825 }
5826 
5827 
5828 static struct bfq_queue *bfq_get_queue(struct bfq_data *bfqd,
5829 				       struct bio *bio, bool is_sync,
5830 				       struct bfq_io_cq *bic,
5831 				       bool respawn)
5832 {
5833 	const int ioprio = IOPRIO_PRIO_LEVEL(bic->ioprio);
5834 	const int ioprio_class = IOPRIO_PRIO_CLASS(bic->ioprio);
5835 	struct bfq_queue **async_bfqq = NULL;
5836 	struct bfq_queue *bfqq;
5837 	struct bfq_group *bfqg;
5838 
5839 	bfqg = bfq_bio_bfqg(bfqd, bio);
5840 	if (!is_sync) {
5841 		async_bfqq = bfq_async_queue_prio(bfqd, bfqg, ioprio_class,
5842 						  ioprio,
5843 						  bfq_actuator_index(bfqd, bio));
5844 		bfqq = *async_bfqq;
5845 		if (bfqq)
5846 			goto out;
5847 	}
5848 
5849 	bfqq = kmem_cache_alloc_node(bfq_pool,
5850 				     GFP_NOWAIT | __GFP_ZERO | __GFP_NOWARN,
5851 				     bfqd->queue->node);
5852 
5853 	if (bfqq) {
5854 		bfq_init_bfqq(bfqd, bfqq, bic, current->pid,
5855 			      is_sync, bfq_actuator_index(bfqd, bio));
5856 		bfq_init_entity(&bfqq->entity, bfqg);
5857 		bfq_log_bfqq(bfqd, bfqq, "allocated");
5858 	} else {
5859 		bfqq = &bfqd->oom_bfqq;
5860 		bfq_log_bfqq(bfqd, bfqq, "using oom bfqq");
5861 		goto out;
5862 	}
5863 
5864 	/*
5865 	 * Pin the queue now that it's allocated, scheduler exit will
5866 	 * prune it.
5867 	 */
5868 	if (async_bfqq) {
5869 		bfqq->ref++; /*
5870 			      * Extra group reference, w.r.t. sync
5871 			      * queue. This extra reference is removed
5872 			      * only if bfqq->bfqg disappears, to
5873 			      * guarantee that this queue is not freed
5874 			      * until its group goes away.
5875 			      */
5876 		bfq_log_bfqq(bfqd, bfqq, "get_queue, bfqq not in async: %p, %d",
5877 			     bfqq, bfqq->ref);
5878 		*async_bfqq = bfqq;
5879 	}
5880 
5881 out:
5882 	bfqq->ref++; /* get a process reference to this queue */
5883 
5884 	if (bfqq != &bfqd->oom_bfqq && is_sync && !respawn)
5885 		bfqq = bfq_do_or_sched_stable_merge(bfqd, bfqq, bic);
5886 	return bfqq;
5887 }
5888 
5889 static void bfq_update_io_thinktime(struct bfq_data *bfqd,
5890 				    struct bfq_queue *bfqq)
5891 {
5892 	struct bfq_ttime *ttime = &bfqq->ttime;
5893 	u64 elapsed;
5894 
5895 	/*
5896 	 * We are really interested in how long it takes for the queue to
5897 	 * become busy when there is no outstanding IO for this queue. So
5898 	 * ignore cases when the bfq queue has already IO queued.
5899 	 */
5900 	if (bfqq->dispatched || bfq_bfqq_busy(bfqq))
5901 		return;
5902 	elapsed = ktime_get_ns() - bfqq->ttime.last_end_request;
5903 	elapsed = min_t(u64, elapsed, 2ULL * bfqd->bfq_slice_idle);
5904 
5905 	ttime->ttime_samples = (7*ttime->ttime_samples + 256) / 8;
5906 	ttime->ttime_total = div_u64(7*ttime->ttime_total + 256*elapsed,  8);
5907 	ttime->ttime_mean = div64_ul(ttime->ttime_total + 128,
5908 				     ttime->ttime_samples);
5909 }
5910 
5911 static void
5912 bfq_update_io_seektime(struct bfq_data *bfqd, struct bfq_queue *bfqq,
5913 		       struct request *rq)
5914 {
5915 	bfqq->seek_history <<= 1;
5916 	bfqq->seek_history |= BFQ_RQ_SEEKY(bfqd, bfqq->last_request_pos, rq);
5917 
5918 	if (bfqq->wr_coeff > 1 &&
5919 	    bfqq->wr_cur_max_time == bfqd->bfq_wr_rt_max_time &&
5920 	    BFQQ_TOTALLY_SEEKY(bfqq)) {
5921 		if (time_is_before_jiffies(bfqq->wr_start_at_switch_to_srt +
5922 					   bfq_wr_duration(bfqd))) {
5923 			/*
5924 			 * In soft_rt weight raising with the
5925 			 * interactive-weight-raising period
5926 			 * elapsed (so no switch back to
5927 			 * interactive weight raising).
5928 			 */
5929 			bfq_bfqq_end_wr(bfqq);
5930 		} else { /*
5931 			  * stopping soft_rt weight raising
5932 			  * while still in interactive period,
5933 			  * switch back to interactive weight
5934 			  * raising
5935 			  */
5936 			switch_back_to_interactive_wr(bfqq, bfqd);
5937 			bfqq->entity.prio_changed = 1;
5938 		}
5939 	}
5940 }
5941 
5942 static void bfq_update_has_short_ttime(struct bfq_data *bfqd,
5943 				       struct bfq_queue *bfqq,
5944 				       struct bfq_io_cq *bic)
5945 {
5946 	bool has_short_ttime = true, state_changed;
5947 
5948 	/*
5949 	 * No need to update has_short_ttime if bfqq is async or in
5950 	 * idle io prio class, or if bfq_slice_idle is zero, because
5951 	 * no device idling is performed for bfqq in this case.
5952 	 */
5953 	if (!bfq_bfqq_sync(bfqq) || bfq_class_idle(bfqq) ||
5954 	    bfqd->bfq_slice_idle == 0)
5955 		return;
5956 
5957 	/* Idle window just restored, statistics are meaningless. */
5958 	if (time_is_after_eq_jiffies(bfqq->split_time +
5959 				     bfqd->bfq_wr_min_idle_time))
5960 		return;
5961 
5962 	/* Think time is infinite if no process is linked to
5963 	 * bfqq. Otherwise check average think time to decide whether
5964 	 * to mark as has_short_ttime. To this goal, compare average
5965 	 * think time with half the I/O-plugging timeout.
5966 	 */
5967 	if (atomic_read(&bic->icq.ioc->active_ref) == 0 ||
5968 	    (bfq_sample_valid(bfqq->ttime.ttime_samples) &&
5969 	     bfqq->ttime.ttime_mean > bfqd->bfq_slice_idle>>1))
5970 		has_short_ttime = false;
5971 
5972 	state_changed = has_short_ttime != bfq_bfqq_has_short_ttime(bfqq);
5973 
5974 	if (has_short_ttime)
5975 		bfq_mark_bfqq_has_short_ttime(bfqq);
5976 	else
5977 		bfq_clear_bfqq_has_short_ttime(bfqq);
5978 
5979 	/*
5980 	 * Until the base value for the total service time gets
5981 	 * finally computed for bfqq, the inject limit does depend on
5982 	 * the think-time state (short|long). In particular, the limit
5983 	 * is 0 or 1 if the think time is deemed, respectively, as
5984 	 * short or long (details in the comments in
5985 	 * bfq_update_inject_limit()). Accordingly, the next
5986 	 * instructions reset the inject limit if the think-time state
5987 	 * has changed and the above base value is still to be
5988 	 * computed.
5989 	 *
5990 	 * However, the reset is performed only if more than 100 ms
5991 	 * have elapsed since the last update of the inject limit, or
5992 	 * (inclusive) if the change is from short to long think
5993 	 * time. The reason for this waiting is as follows.
5994 	 *
5995 	 * bfqq may have a long think time because of a
5996 	 * synchronization with some other queue, i.e., because the
5997 	 * I/O of some other queue may need to be completed for bfqq
5998 	 * to receive new I/O. Details in the comments on the choice
5999 	 * of the queue for injection in bfq_select_queue().
6000 	 *
6001 	 * As stressed in those comments, if such a synchronization is
6002 	 * actually in place, then, without injection on bfqq, the
6003 	 * blocking I/O cannot happen to served while bfqq is in
6004 	 * service. As a consequence, if bfqq is granted
6005 	 * I/O-dispatch-plugging, then bfqq remains empty, and no I/O
6006 	 * is dispatched, until the idle timeout fires. This is likely
6007 	 * to result in lower bandwidth and higher latencies for bfqq,
6008 	 * and in a severe loss of total throughput.
6009 	 *
6010 	 * On the opposite end, a non-zero inject limit may allow the
6011 	 * I/O that blocks bfqq to be executed soon, and therefore
6012 	 * bfqq to receive new I/O soon.
6013 	 *
6014 	 * But, if the blocking gets actually eliminated, then the
6015 	 * next think-time sample for bfqq may be very low. This in
6016 	 * turn may cause bfqq's think time to be deemed
6017 	 * short. Without the 100 ms barrier, this new state change
6018 	 * would cause the body of the next if to be executed
6019 	 * immediately. But this would set to 0 the inject
6020 	 * limit. Without injection, the blocking I/O would cause the
6021 	 * think time of bfqq to become long again, and therefore the
6022 	 * inject limit to be raised again, and so on. The only effect
6023 	 * of such a steady oscillation between the two think-time
6024 	 * states would be to prevent effective injection on bfqq.
6025 	 *
6026 	 * In contrast, if the inject limit is not reset during such a
6027 	 * long time interval as 100 ms, then the number of short
6028 	 * think time samples can grow significantly before the reset
6029 	 * is performed. As a consequence, the think time state can
6030 	 * become stable before the reset. Therefore there will be no
6031 	 * state change when the 100 ms elapse, and no reset of the
6032 	 * inject limit. The inject limit remains steadily equal to 1
6033 	 * both during and after the 100 ms. So injection can be
6034 	 * performed at all times, and throughput gets boosted.
6035 	 *
6036 	 * An inject limit equal to 1 is however in conflict, in
6037 	 * general, with the fact that the think time of bfqq is
6038 	 * short, because injection may be likely to delay bfqq's I/O
6039 	 * (as explained in the comments in
6040 	 * bfq_update_inject_limit()). But this does not happen in
6041 	 * this special case, because bfqq's low think time is due to
6042 	 * an effective handling of a synchronization, through
6043 	 * injection. In this special case, bfqq's I/O does not get
6044 	 * delayed by injection; on the contrary, bfqq's I/O is
6045 	 * brought forward, because it is not blocked for
6046 	 * milliseconds.
6047 	 *
6048 	 * In addition, serving the blocking I/O much sooner, and much
6049 	 * more frequently than once per I/O-plugging timeout, makes
6050 	 * it much quicker to detect a waker queue (the concept of
6051 	 * waker queue is defined in the comments in
6052 	 * bfq_add_request()). This makes it possible to start sooner
6053 	 * to boost throughput more effectively, by injecting the I/O
6054 	 * of the waker queue unconditionally on every
6055 	 * bfq_dispatch_request().
6056 	 *
6057 	 * One last, important benefit of not resetting the inject
6058 	 * limit before 100 ms is that, during this time interval, the
6059 	 * base value for the total service time is likely to get
6060 	 * finally computed for bfqq, freeing the inject limit from
6061 	 * its relation with the think time.
6062 	 */
6063 	if (state_changed && bfqq->last_serv_time_ns == 0 &&
6064 	    (time_is_before_eq_jiffies(bfqq->decrease_time_jif +
6065 				      msecs_to_jiffies(100)) ||
6066 	     !has_short_ttime))
6067 		bfq_reset_inject_limit(bfqd, bfqq);
6068 }
6069 
6070 /*
6071  * Called when a new fs request (rq) is added to bfqq.  Check if there's
6072  * something we should do about it.
6073  */
6074 static void bfq_rq_enqueued(struct bfq_data *bfqd, struct bfq_queue *bfqq,
6075 			    struct request *rq)
6076 {
6077 	if (rq->cmd_flags & REQ_META)
6078 		bfqq->meta_pending++;
6079 
6080 	bfqq->last_request_pos = blk_rq_pos(rq) + blk_rq_sectors(rq);
6081 
6082 	if (bfqq == bfqd->in_service_queue && bfq_bfqq_wait_request(bfqq)) {
6083 		bool small_req = bfqq->queued[rq_is_sync(rq)] == 1 &&
6084 				 blk_rq_sectors(rq) < 32;
6085 		bool budget_timeout = bfq_bfqq_budget_timeout(bfqq);
6086 
6087 		/*
6088 		 * There is just this request queued: if
6089 		 * - the request is small, and
6090 		 * - we are idling to boost throughput, and
6091 		 * - the queue is not to be expired,
6092 		 * then just exit.
6093 		 *
6094 		 * In this way, if the device is being idled to wait
6095 		 * for a new request from the in-service queue, we
6096 		 * avoid unplugging the device and committing the
6097 		 * device to serve just a small request. In contrast
6098 		 * we wait for the block layer to decide when to
6099 		 * unplug the device: hopefully, new requests will be
6100 		 * merged to this one quickly, then the device will be
6101 		 * unplugged and larger requests will be dispatched.
6102 		 */
6103 		if (small_req && idling_boosts_thr_without_issues(bfqd, bfqq) &&
6104 		    !budget_timeout)
6105 			return;
6106 
6107 		/*
6108 		 * A large enough request arrived, or idling is being
6109 		 * performed to preserve service guarantees, or
6110 		 * finally the queue is to be expired: in all these
6111 		 * cases disk idling is to be stopped, so clear
6112 		 * wait_request flag and reset timer.
6113 		 */
6114 		bfq_clear_bfqq_wait_request(bfqq);
6115 		hrtimer_try_to_cancel(&bfqd->idle_slice_timer);
6116 
6117 		/*
6118 		 * The queue is not empty, because a new request just
6119 		 * arrived. Hence we can safely expire the queue, in
6120 		 * case of budget timeout, without risking that the
6121 		 * timestamps of the queue are not updated correctly.
6122 		 * See [1] for more details.
6123 		 */
6124 		if (budget_timeout)
6125 			bfq_bfqq_expire(bfqd, bfqq, false,
6126 					BFQQE_BUDGET_TIMEOUT);
6127 	}
6128 }
6129 
6130 static void bfqq_request_allocated(struct bfq_queue *bfqq)
6131 {
6132 	struct bfq_entity *entity = &bfqq->entity;
6133 
6134 	for_each_entity(entity)
6135 		entity->allocated++;
6136 }
6137 
6138 static void bfqq_request_freed(struct bfq_queue *bfqq)
6139 {
6140 	struct bfq_entity *entity = &bfqq->entity;
6141 
6142 	for_each_entity(entity)
6143 		entity->allocated--;
6144 }
6145 
6146 /* returns true if it causes the idle timer to be disabled */
6147 static bool __bfq_insert_request(struct bfq_data *bfqd, struct request *rq)
6148 {
6149 	struct bfq_queue *bfqq = RQ_BFQQ(rq),
6150 		*new_bfqq = bfq_setup_cooperator(bfqd, bfqq, rq, true,
6151 						 RQ_BIC(rq));
6152 	bool waiting, idle_timer_disabled = false;
6153 
6154 	if (new_bfqq) {
6155 		/*
6156 		 * Release the request's reference to the old bfqq
6157 		 * and make sure one is taken to the shared queue.
6158 		 */
6159 		bfqq_request_allocated(new_bfqq);
6160 		bfqq_request_freed(bfqq);
6161 		new_bfqq->ref++;
6162 		/*
6163 		 * If the bic associated with the process
6164 		 * issuing this request still points to bfqq
6165 		 * (and thus has not been already redirected
6166 		 * to new_bfqq or even some other bfq_queue),
6167 		 * then complete the merge and redirect it to
6168 		 * new_bfqq.
6169 		 */
6170 		if (bic_to_bfqq(RQ_BIC(rq), true,
6171 				bfq_actuator_index(bfqd, rq->bio)) == bfqq)
6172 			bfq_merge_bfqqs(bfqd, RQ_BIC(rq),
6173 					bfqq, new_bfqq);
6174 
6175 		bfq_clear_bfqq_just_created(bfqq);
6176 		/*
6177 		 * rq is about to be enqueued into new_bfqq,
6178 		 * release rq reference on bfqq
6179 		 */
6180 		bfq_put_queue(bfqq);
6181 		rq->elv.priv[1] = new_bfqq;
6182 		bfqq = new_bfqq;
6183 	}
6184 
6185 	bfq_update_io_thinktime(bfqd, bfqq);
6186 	bfq_update_has_short_ttime(bfqd, bfqq, RQ_BIC(rq));
6187 	bfq_update_io_seektime(bfqd, bfqq, rq);
6188 
6189 	waiting = bfqq && bfq_bfqq_wait_request(bfqq);
6190 	bfq_add_request(rq);
6191 	idle_timer_disabled = waiting && !bfq_bfqq_wait_request(bfqq);
6192 
6193 	rq->fifo_time = ktime_get_ns() + bfqd->bfq_fifo_expire[rq_is_sync(rq)];
6194 	list_add_tail(&rq->queuelist, &bfqq->fifo);
6195 
6196 	bfq_rq_enqueued(bfqd, bfqq, rq);
6197 
6198 	return idle_timer_disabled;
6199 }
6200 
6201 #ifdef CONFIG_BFQ_CGROUP_DEBUG
6202 static void bfq_update_insert_stats(struct request_queue *q,
6203 				    struct bfq_queue *bfqq,
6204 				    bool idle_timer_disabled,
6205 				    blk_opf_t cmd_flags)
6206 {
6207 	if (!bfqq)
6208 		return;
6209 
6210 	/*
6211 	 * bfqq still exists, because it can disappear only after
6212 	 * either it is merged with another queue, or the process it
6213 	 * is associated with exits. But both actions must be taken by
6214 	 * the same process currently executing this flow of
6215 	 * instructions.
6216 	 *
6217 	 * In addition, the following queue lock guarantees that
6218 	 * bfqq_group(bfqq) exists as well.
6219 	 */
6220 	spin_lock_irq(&q->queue_lock);
6221 	bfqg_stats_update_io_add(bfqq_group(bfqq), bfqq, cmd_flags);
6222 	if (idle_timer_disabled)
6223 		bfqg_stats_update_idle_time(bfqq_group(bfqq));
6224 	spin_unlock_irq(&q->queue_lock);
6225 }
6226 #else
6227 static inline void bfq_update_insert_stats(struct request_queue *q,
6228 					   struct bfq_queue *bfqq,
6229 					   bool idle_timer_disabled,
6230 					   blk_opf_t cmd_flags) {}
6231 #endif /* CONFIG_BFQ_CGROUP_DEBUG */
6232 
6233 static struct bfq_queue *bfq_init_rq(struct request *rq);
6234 
6235 static void bfq_insert_request(struct blk_mq_hw_ctx *hctx, struct request *rq,
6236 			       blk_insert_t flags)
6237 {
6238 	struct request_queue *q = hctx->queue;
6239 	struct bfq_data *bfqd = q->elevator->elevator_data;
6240 	struct bfq_queue *bfqq;
6241 	bool idle_timer_disabled = false;
6242 	blk_opf_t cmd_flags;
6243 	LIST_HEAD(free);
6244 
6245 #ifdef CONFIG_BFQ_GROUP_IOSCHED
6246 	if (!cgroup_subsys_on_dfl(io_cgrp_subsys) && rq->bio)
6247 		bfqg_stats_update_legacy_io(q, rq);
6248 #endif
6249 	spin_lock_irq(&bfqd->lock);
6250 	bfqq = bfq_init_rq(rq);
6251 	if (blk_mq_sched_try_insert_merge(q, rq, &free)) {
6252 		spin_unlock_irq(&bfqd->lock);
6253 		blk_mq_free_requests(&free);
6254 		return;
6255 	}
6256 
6257 	trace_block_rq_insert(rq);
6258 
6259 	if (flags & BLK_MQ_INSERT_AT_HEAD) {
6260 		list_add(&rq->queuelist, &bfqd->dispatch);
6261 	} else if (!bfqq) {
6262 		list_add_tail(&rq->queuelist, &bfqd->dispatch);
6263 	} else {
6264 		idle_timer_disabled = __bfq_insert_request(bfqd, rq);
6265 		/*
6266 		 * Update bfqq, because, if a queue merge has occurred
6267 		 * in __bfq_insert_request, then rq has been
6268 		 * redirected into a new queue.
6269 		 */
6270 		bfqq = RQ_BFQQ(rq);
6271 
6272 		if (rq_mergeable(rq)) {
6273 			elv_rqhash_add(q, rq);
6274 			if (!q->last_merge)
6275 				q->last_merge = rq;
6276 		}
6277 	}
6278 
6279 	/*
6280 	 * Cache cmd_flags before releasing scheduler lock, because rq
6281 	 * may disappear afterwards (for example, because of a request
6282 	 * merge).
6283 	 */
6284 	cmd_flags = rq->cmd_flags;
6285 	spin_unlock_irq(&bfqd->lock);
6286 
6287 	bfq_update_insert_stats(q, bfqq, idle_timer_disabled,
6288 				cmd_flags);
6289 }
6290 
6291 static void bfq_insert_requests(struct blk_mq_hw_ctx *hctx,
6292 				struct list_head *list,
6293 				blk_insert_t flags)
6294 {
6295 	while (!list_empty(list)) {
6296 		struct request *rq;
6297 
6298 		rq = list_first_entry(list, struct request, queuelist);
6299 		list_del_init(&rq->queuelist);
6300 		bfq_insert_request(hctx, rq, flags);
6301 	}
6302 }
6303 
6304 static void bfq_update_hw_tag(struct bfq_data *bfqd)
6305 {
6306 	struct bfq_queue *bfqq = bfqd->in_service_queue;
6307 
6308 	bfqd->max_rq_in_driver = max_t(int, bfqd->max_rq_in_driver,
6309 				       bfqd->tot_rq_in_driver);
6310 
6311 	if (bfqd->hw_tag == 1)
6312 		return;
6313 
6314 	/*
6315 	 * This sample is valid if the number of outstanding requests
6316 	 * is large enough to allow a queueing behavior.  Note that the
6317 	 * sum is not exact, as it's not taking into account deactivated
6318 	 * requests.
6319 	 */
6320 	if (bfqd->tot_rq_in_driver + bfqd->queued <= BFQ_HW_QUEUE_THRESHOLD)
6321 		return;
6322 
6323 	/*
6324 	 * If active queue hasn't enough requests and can idle, bfq might not
6325 	 * dispatch sufficient requests to hardware. Don't zero hw_tag in this
6326 	 * case
6327 	 */
6328 	if (bfqq && bfq_bfqq_has_short_ttime(bfqq) &&
6329 	    bfqq->dispatched + bfqq->queued[0] + bfqq->queued[1] <
6330 	    BFQ_HW_QUEUE_THRESHOLD &&
6331 	    bfqd->tot_rq_in_driver < BFQ_HW_QUEUE_THRESHOLD)
6332 		return;
6333 
6334 	if (bfqd->hw_tag_samples++ < BFQ_HW_QUEUE_SAMPLES)
6335 		return;
6336 
6337 	bfqd->hw_tag = bfqd->max_rq_in_driver > BFQ_HW_QUEUE_THRESHOLD;
6338 	bfqd->max_rq_in_driver = 0;
6339 	bfqd->hw_tag_samples = 0;
6340 
6341 	bfqd->nonrot_with_queueing =
6342 		blk_queue_nonrot(bfqd->queue) && bfqd->hw_tag;
6343 }
6344 
6345 static void bfq_completed_request(struct bfq_queue *bfqq, struct bfq_data *bfqd)
6346 {
6347 	u64 now_ns;
6348 	u32 delta_us;
6349 
6350 	bfq_update_hw_tag(bfqd);
6351 
6352 	bfqd->rq_in_driver[bfqq->actuator_idx]--;
6353 	bfqd->tot_rq_in_driver--;
6354 	bfqq->dispatched--;
6355 
6356 	if (!bfqq->dispatched && !bfq_bfqq_busy(bfqq)) {
6357 		/*
6358 		 * Set budget_timeout (which we overload to store the
6359 		 * time at which the queue remains with no backlog and
6360 		 * no outstanding request; used by the weight-raising
6361 		 * mechanism).
6362 		 */
6363 		bfqq->budget_timeout = jiffies;
6364 
6365 		bfq_del_bfqq_in_groups_with_pending_reqs(bfqq);
6366 		bfq_weights_tree_remove(bfqq);
6367 	}
6368 
6369 	now_ns = ktime_get_ns();
6370 
6371 	bfqq->ttime.last_end_request = now_ns;
6372 
6373 	/*
6374 	 * Using us instead of ns, to get a reasonable precision in
6375 	 * computing rate in next check.
6376 	 */
6377 	delta_us = div_u64(now_ns - bfqd->last_completion, NSEC_PER_USEC);
6378 
6379 	/*
6380 	 * If the request took rather long to complete, and, according
6381 	 * to the maximum request size recorded, this completion latency
6382 	 * implies that the request was certainly served at a very low
6383 	 * rate (less than 1M sectors/sec), then the whole observation
6384 	 * interval that lasts up to this time instant cannot be a
6385 	 * valid time interval for computing a new peak rate.  Invoke
6386 	 * bfq_update_rate_reset to have the following three steps
6387 	 * taken:
6388 	 * - close the observation interval at the last (previous)
6389 	 *   request dispatch or completion
6390 	 * - compute rate, if possible, for that observation interval
6391 	 * - reset to zero samples, which will trigger a proper
6392 	 *   re-initialization of the observation interval on next
6393 	 *   dispatch
6394 	 */
6395 	if (delta_us > BFQ_MIN_TT/NSEC_PER_USEC &&
6396 	   (bfqd->last_rq_max_size<<BFQ_RATE_SHIFT)/delta_us <
6397 			1UL<<(BFQ_RATE_SHIFT - 10))
6398 		bfq_update_rate_reset(bfqd, NULL);
6399 	bfqd->last_completion = now_ns;
6400 	/*
6401 	 * Shared queues are likely to receive I/O at a high
6402 	 * rate. This may deceptively let them be considered as wakers
6403 	 * of other queues. But a false waker will unjustly steal
6404 	 * bandwidth to its supposedly woken queue. So considering
6405 	 * also shared queues in the waking mechanism may cause more
6406 	 * control troubles than throughput benefits. Then reset
6407 	 * last_completed_rq_bfqq if bfqq is a shared queue.
6408 	 */
6409 	if (!bfq_bfqq_coop(bfqq))
6410 		bfqd->last_completed_rq_bfqq = bfqq;
6411 	else
6412 		bfqd->last_completed_rq_bfqq = NULL;
6413 
6414 	/*
6415 	 * If we are waiting to discover whether the request pattern
6416 	 * of the task associated with the queue is actually
6417 	 * isochronous, and both requisites for this condition to hold
6418 	 * are now satisfied, then compute soft_rt_next_start (see the
6419 	 * comments on the function bfq_bfqq_softrt_next_start()). We
6420 	 * do not compute soft_rt_next_start if bfqq is in interactive
6421 	 * weight raising (see the comments in bfq_bfqq_expire() for
6422 	 * an explanation). We schedule this delayed update when bfqq
6423 	 * expires, if it still has in-flight requests.
6424 	 */
6425 	if (bfq_bfqq_softrt_update(bfqq) && bfqq->dispatched == 0 &&
6426 	    RB_EMPTY_ROOT(&bfqq->sort_list) &&
6427 	    bfqq->wr_coeff != bfqd->bfq_wr_coeff)
6428 		bfqq->soft_rt_next_start =
6429 			bfq_bfqq_softrt_next_start(bfqd, bfqq);
6430 
6431 	/*
6432 	 * If this is the in-service queue, check if it needs to be expired,
6433 	 * or if we want to idle in case it has no pending requests.
6434 	 */
6435 	if (bfqd->in_service_queue == bfqq) {
6436 		if (bfq_bfqq_must_idle(bfqq)) {
6437 			if (bfqq->dispatched == 0)
6438 				bfq_arm_slice_timer(bfqd);
6439 			/*
6440 			 * If we get here, we do not expire bfqq, even
6441 			 * if bfqq was in budget timeout or had no
6442 			 * more requests (as controlled in the next
6443 			 * conditional instructions). The reason for
6444 			 * not expiring bfqq is as follows.
6445 			 *
6446 			 * Here bfqq->dispatched > 0 holds, but
6447 			 * bfq_bfqq_must_idle() returned true. This
6448 			 * implies that, even if no request arrives
6449 			 * for bfqq before bfqq->dispatched reaches 0,
6450 			 * bfqq will, however, not be expired on the
6451 			 * completion event that causes bfqq->dispatch
6452 			 * to reach zero. In contrast, on this event,
6453 			 * bfqq will start enjoying device idling
6454 			 * (I/O-dispatch plugging).
6455 			 *
6456 			 * But, if we expired bfqq here, bfqq would
6457 			 * not have the chance to enjoy device idling
6458 			 * when bfqq->dispatched finally reaches
6459 			 * zero. This would expose bfqq to violation
6460 			 * of its reserved service guarantees.
6461 			 */
6462 			return;
6463 		} else if (bfq_may_expire_for_budg_timeout(bfqq))
6464 			bfq_bfqq_expire(bfqd, bfqq, false,
6465 					BFQQE_BUDGET_TIMEOUT);
6466 		else if (RB_EMPTY_ROOT(&bfqq->sort_list) &&
6467 			 (bfqq->dispatched == 0 ||
6468 			  !bfq_better_to_idle(bfqq)))
6469 			bfq_bfqq_expire(bfqd, bfqq, false,
6470 					BFQQE_NO_MORE_REQUESTS);
6471 	}
6472 
6473 	if (!bfqd->tot_rq_in_driver)
6474 		bfq_schedule_dispatch(bfqd);
6475 }
6476 
6477 /*
6478  * The processes associated with bfqq may happen to generate their
6479  * cumulative I/O at a lower rate than the rate at which the device
6480  * could serve the same I/O. This is rather probable, e.g., if only
6481  * one process is associated with bfqq and the device is an SSD. It
6482  * results in bfqq becoming often empty while in service. In this
6483  * respect, if BFQ is allowed to switch to another queue when bfqq
6484  * remains empty, then the device goes on being fed with I/O requests,
6485  * and the throughput is not affected. In contrast, if BFQ is not
6486  * allowed to switch to another queue---because bfqq is sync and
6487  * I/O-dispatch needs to be plugged while bfqq is temporarily
6488  * empty---then, during the service of bfqq, there will be frequent
6489  * "service holes", i.e., time intervals during which bfqq gets empty
6490  * and the device can only consume the I/O already queued in its
6491  * hardware queues. During service holes, the device may even get to
6492  * remaining idle. In the end, during the service of bfqq, the device
6493  * is driven at a lower speed than the one it can reach with the kind
6494  * of I/O flowing through bfqq.
6495  *
6496  * To counter this loss of throughput, BFQ implements a "request
6497  * injection mechanism", which tries to fill the above service holes
6498  * with I/O requests taken from other queues. The hard part in this
6499  * mechanism is finding the right amount of I/O to inject, so as to
6500  * both boost throughput and not break bfqq's bandwidth and latency
6501  * guarantees. In this respect, the mechanism maintains a per-queue
6502  * inject limit, computed as below. While bfqq is empty, the injection
6503  * mechanism dispatches extra I/O requests only until the total number
6504  * of I/O requests in flight---i.e., already dispatched but not yet
6505  * completed---remains lower than this limit.
6506  *
6507  * A first definition comes in handy to introduce the algorithm by
6508  * which the inject limit is computed.  We define as first request for
6509  * bfqq, an I/O request for bfqq that arrives while bfqq is in
6510  * service, and causes bfqq to switch from empty to non-empty. The
6511  * algorithm updates the limit as a function of the effect of
6512  * injection on the service times of only the first requests of
6513  * bfqq. The reason for this restriction is that these are the
6514  * requests whose service time is affected most, because they are the
6515  * first to arrive after injection possibly occurred.
6516  *
6517  * To evaluate the effect of injection, the algorithm measures the
6518  * "total service time" of first requests. We define as total service
6519  * time of an I/O request, the time that elapses since when the
6520  * request is enqueued into bfqq, to when it is completed. This
6521  * quantity allows the whole effect of injection to be measured. It is
6522  * easy to see why. Suppose that some requests of other queues are
6523  * actually injected while bfqq is empty, and that a new request R
6524  * then arrives for bfqq. If the device does start to serve all or
6525  * part of the injected requests during the service hole, then,
6526  * because of this extra service, it may delay the next invocation of
6527  * the dispatch hook of BFQ. Then, even after R gets eventually
6528  * dispatched, the device may delay the actual service of R if it is
6529  * still busy serving the extra requests, or if it decides to serve,
6530  * before R, some extra request still present in its queues. As a
6531  * conclusion, the cumulative extra delay caused by injection can be
6532  * easily evaluated by just comparing the total service time of first
6533  * requests with and without injection.
6534  *
6535  * The limit-update algorithm works as follows. On the arrival of a
6536  * first request of bfqq, the algorithm measures the total time of the
6537  * request only if one of the three cases below holds, and, for each
6538  * case, it updates the limit as described below:
6539  *
6540  * (1) If there is no in-flight request. This gives a baseline for the
6541  *     total service time of the requests of bfqq. If the baseline has
6542  *     not been computed yet, then, after computing it, the limit is
6543  *     set to 1, to start boosting throughput, and to prepare the
6544  *     ground for the next case. If the baseline has already been
6545  *     computed, then it is updated, in case it results to be lower
6546  *     than the previous value.
6547  *
6548  * (2) If the limit is higher than 0 and there are in-flight
6549  *     requests. By comparing the total service time in this case with
6550  *     the above baseline, it is possible to know at which extent the
6551  *     current value of the limit is inflating the total service
6552  *     time. If the inflation is below a certain threshold, then bfqq
6553  *     is assumed to be suffering from no perceivable loss of its
6554  *     service guarantees, and the limit is even tentatively
6555  *     increased. If the inflation is above the threshold, then the
6556  *     limit is decreased. Due to the lack of any hysteresis, this
6557  *     logic makes the limit oscillate even in steady workload
6558  *     conditions. Yet we opted for it, because it is fast in reaching
6559  *     the best value for the limit, as a function of the current I/O
6560  *     workload. To reduce oscillations, this step is disabled for a
6561  *     short time interval after the limit happens to be decreased.
6562  *
6563  * (3) Periodically, after resetting the limit, to make sure that the
6564  *     limit eventually drops in case the workload changes. This is
6565  *     needed because, after the limit has gone safely up for a
6566  *     certain workload, it is impossible to guess whether the
6567  *     baseline total service time may have changed, without measuring
6568  *     it again without injection. A more effective version of this
6569  *     step might be to just sample the baseline, by interrupting
6570  *     injection only once, and then to reset/lower the limit only if
6571  *     the total service time with the current limit does happen to be
6572  *     too large.
6573  *
6574  * More details on each step are provided in the comments on the
6575  * pieces of code that implement these steps: the branch handling the
6576  * transition from empty to non empty in bfq_add_request(), the branch
6577  * handling injection in bfq_select_queue(), and the function
6578  * bfq_choose_bfqq_for_injection(). These comments also explain some
6579  * exceptions, made by the injection mechanism in some special cases.
6580  */
6581 static void bfq_update_inject_limit(struct bfq_data *bfqd,
6582 				    struct bfq_queue *bfqq)
6583 {
6584 	u64 tot_time_ns = ktime_get_ns() - bfqd->last_empty_occupied_ns;
6585 	unsigned int old_limit = bfqq->inject_limit;
6586 
6587 	if (bfqq->last_serv_time_ns > 0 && bfqd->rqs_injected) {
6588 		u64 threshold = (bfqq->last_serv_time_ns * 3)>>1;
6589 
6590 		if (tot_time_ns >= threshold && old_limit > 0) {
6591 			bfqq->inject_limit--;
6592 			bfqq->decrease_time_jif = jiffies;
6593 		} else if (tot_time_ns < threshold &&
6594 			   old_limit <= bfqd->max_rq_in_driver)
6595 			bfqq->inject_limit++;
6596 	}
6597 
6598 	/*
6599 	 * Either we still have to compute the base value for the
6600 	 * total service time, and there seem to be the right
6601 	 * conditions to do it, or we can lower the last base value
6602 	 * computed.
6603 	 *
6604 	 * NOTE: (bfqd->tot_rq_in_driver == 1) means that there is no I/O
6605 	 * request in flight, because this function is in the code
6606 	 * path that handles the completion of a request of bfqq, and,
6607 	 * in particular, this function is executed before
6608 	 * bfqd->tot_rq_in_driver is decremented in such a code path.
6609 	 */
6610 	if ((bfqq->last_serv_time_ns == 0 && bfqd->tot_rq_in_driver == 1) ||
6611 	    tot_time_ns < bfqq->last_serv_time_ns) {
6612 		if (bfqq->last_serv_time_ns == 0) {
6613 			/*
6614 			 * Now we certainly have a base value: make sure we
6615 			 * start trying injection.
6616 			 */
6617 			bfqq->inject_limit = max_t(unsigned int, 1, old_limit);
6618 		}
6619 		bfqq->last_serv_time_ns = tot_time_ns;
6620 	} else if (!bfqd->rqs_injected && bfqd->tot_rq_in_driver == 1)
6621 		/*
6622 		 * No I/O injected and no request still in service in
6623 		 * the drive: these are the exact conditions for
6624 		 * computing the base value of the total service time
6625 		 * for bfqq. So let's update this value, because it is
6626 		 * rather variable. For example, it varies if the size
6627 		 * or the spatial locality of the I/O requests in bfqq
6628 		 * change.
6629 		 */
6630 		bfqq->last_serv_time_ns = tot_time_ns;
6631 
6632 
6633 	/* update complete, not waiting for any request completion any longer */
6634 	bfqd->waited_rq = NULL;
6635 	bfqd->rqs_injected = false;
6636 }
6637 
6638 /*
6639  * Handle either a requeue or a finish for rq. The things to do are
6640  * the same in both cases: all references to rq are to be dropped. In
6641  * particular, rq is considered completed from the point of view of
6642  * the scheduler.
6643  */
6644 static void bfq_finish_requeue_request(struct request *rq)
6645 {
6646 	struct bfq_queue *bfqq = RQ_BFQQ(rq);
6647 	struct bfq_data *bfqd;
6648 	unsigned long flags;
6649 
6650 	/*
6651 	 * rq either is not associated with any icq, or is an already
6652 	 * requeued request that has not (yet) been re-inserted into
6653 	 * a bfq_queue.
6654 	 */
6655 	if (!rq->elv.icq || !bfqq)
6656 		return;
6657 
6658 	bfqd = bfqq->bfqd;
6659 
6660 	if (rq->rq_flags & RQF_STARTED)
6661 		bfqg_stats_update_completion(bfqq_group(bfqq),
6662 					     rq->start_time_ns,
6663 					     rq->io_start_time_ns,
6664 					     rq->cmd_flags);
6665 
6666 	spin_lock_irqsave(&bfqd->lock, flags);
6667 	if (likely(rq->rq_flags & RQF_STARTED)) {
6668 		if (rq == bfqd->waited_rq)
6669 			bfq_update_inject_limit(bfqd, bfqq);
6670 
6671 		bfq_completed_request(bfqq, bfqd);
6672 	}
6673 	bfqq_request_freed(bfqq);
6674 	bfq_put_queue(bfqq);
6675 	RQ_BIC(rq)->requests--;
6676 	spin_unlock_irqrestore(&bfqd->lock, flags);
6677 
6678 	/*
6679 	 * Reset private fields. In case of a requeue, this allows
6680 	 * this function to correctly do nothing if it is spuriously
6681 	 * invoked again on this same request (see the check at the
6682 	 * beginning of the function). Probably, a better general
6683 	 * design would be to prevent blk-mq from invoking the requeue
6684 	 * or finish hooks of an elevator, for a request that is not
6685 	 * referred by that elevator.
6686 	 *
6687 	 * Resetting the following fields would break the
6688 	 * request-insertion logic if rq is re-inserted into a bfq
6689 	 * internal queue, without a re-preparation. Here we assume
6690 	 * that re-insertions of requeued requests, without
6691 	 * re-preparation, can happen only for pass_through or at_head
6692 	 * requests (which are not re-inserted into bfq internal
6693 	 * queues).
6694 	 */
6695 	rq->elv.priv[0] = NULL;
6696 	rq->elv.priv[1] = NULL;
6697 }
6698 
6699 static void bfq_finish_request(struct request *rq)
6700 {
6701 	bfq_finish_requeue_request(rq);
6702 
6703 	if (rq->elv.icq) {
6704 		put_io_context(rq->elv.icq->ioc);
6705 		rq->elv.icq = NULL;
6706 	}
6707 }
6708 
6709 /*
6710  * Removes the association between the current task and bfqq, assuming
6711  * that bic points to the bfq iocontext of the task.
6712  * Returns NULL if a new bfqq should be allocated, or the old bfqq if this
6713  * was the last process referring to that bfqq.
6714  */
6715 static struct bfq_queue *
6716 bfq_split_bfqq(struct bfq_io_cq *bic, struct bfq_queue *bfqq)
6717 {
6718 	bfq_log_bfqq(bfqq->bfqd, bfqq, "splitting queue");
6719 
6720 	if (bfqq_process_refs(bfqq) == 1) {
6721 		bfqq->pid = current->pid;
6722 		bfq_clear_bfqq_coop(bfqq);
6723 		bfq_clear_bfqq_split_coop(bfqq);
6724 		return bfqq;
6725 	}
6726 
6727 	bic_set_bfqq(bic, NULL, true, bfqq->actuator_idx);
6728 
6729 	bfq_put_cooperator(bfqq);
6730 
6731 	bfq_release_process_ref(bfqq->bfqd, bfqq);
6732 	return NULL;
6733 }
6734 
6735 static struct bfq_queue *bfq_get_bfqq_handle_split(struct bfq_data *bfqd,
6736 						   struct bfq_io_cq *bic,
6737 						   struct bio *bio,
6738 						   bool split, bool is_sync,
6739 						   bool *new_queue)
6740 {
6741 	unsigned int act_idx = bfq_actuator_index(bfqd, bio);
6742 	struct bfq_queue *bfqq = bic_to_bfqq(bic, is_sync, act_idx);
6743 	struct bfq_iocq_bfqq_data *bfqq_data = &bic->bfqq_data[act_idx];
6744 
6745 	if (likely(bfqq && bfqq != &bfqd->oom_bfqq))
6746 		return bfqq;
6747 
6748 	if (new_queue)
6749 		*new_queue = true;
6750 
6751 	if (bfqq)
6752 		bfq_put_queue(bfqq);
6753 	bfqq = bfq_get_queue(bfqd, bio, is_sync, bic, split);
6754 
6755 	bic_set_bfqq(bic, bfqq, is_sync, act_idx);
6756 	if (split && is_sync) {
6757 		if ((bfqq_data->was_in_burst_list && bfqd->large_burst) ||
6758 		    bfqq_data->saved_in_large_burst)
6759 			bfq_mark_bfqq_in_large_burst(bfqq);
6760 		else {
6761 			bfq_clear_bfqq_in_large_burst(bfqq);
6762 			if (bfqq_data->was_in_burst_list)
6763 				/*
6764 				 * If bfqq was in the current
6765 				 * burst list before being
6766 				 * merged, then we have to add
6767 				 * it back. And we do not need
6768 				 * to increase burst_size, as
6769 				 * we did not decrement
6770 				 * burst_size when we removed
6771 				 * bfqq from the burst list as
6772 				 * a consequence of a merge
6773 				 * (see comments in
6774 				 * bfq_put_queue). In this
6775 				 * respect, it would be rather
6776 				 * costly to know whether the
6777 				 * current burst list is still
6778 				 * the same burst list from
6779 				 * which bfqq was removed on
6780 				 * the merge. To avoid this
6781 				 * cost, if bfqq was in a
6782 				 * burst list, then we add
6783 				 * bfqq to the current burst
6784 				 * list without any further
6785 				 * check. This can cause
6786 				 * inappropriate insertions,
6787 				 * but rarely enough to not
6788 				 * harm the detection of large
6789 				 * bursts significantly.
6790 				 */
6791 				hlist_add_head(&bfqq->burst_list_node,
6792 					       &bfqd->burst_list);
6793 		}
6794 		bfqq->split_time = jiffies;
6795 	}
6796 
6797 	return bfqq;
6798 }
6799 
6800 /*
6801  * Only reset private fields. The actual request preparation will be
6802  * performed by bfq_init_rq, when rq is either inserted or merged. See
6803  * comments on bfq_init_rq for the reason behind this delayed
6804  * preparation.
6805  */
6806 static void bfq_prepare_request(struct request *rq)
6807 {
6808 	rq->elv.icq = ioc_find_get_icq(rq->q);
6809 
6810 	/*
6811 	 * Regardless of whether we have an icq attached, we have to
6812 	 * clear the scheduler pointers, as they might point to
6813 	 * previously allocated bic/bfqq structs.
6814 	 */
6815 	rq->elv.priv[0] = rq->elv.priv[1] = NULL;
6816 }
6817 
6818 /*
6819  * If needed, init rq, allocate bfq data structures associated with
6820  * rq, and increment reference counters in the destination bfq_queue
6821  * for rq. Return the destination bfq_queue for rq, or NULL is rq is
6822  * not associated with any bfq_queue.
6823  *
6824  * This function is invoked by the functions that perform rq insertion
6825  * or merging. One may have expected the above preparation operations
6826  * to be performed in bfq_prepare_request, and not delayed to when rq
6827  * is inserted or merged. The rationale behind this delayed
6828  * preparation is that, after the prepare_request hook is invoked for
6829  * rq, rq may still be transformed into a request with no icq, i.e., a
6830  * request not associated with any queue. No bfq hook is invoked to
6831  * signal this transformation. As a consequence, should these
6832  * preparation operations be performed when the prepare_request hook
6833  * is invoked, and should rq be transformed one moment later, bfq
6834  * would end up in an inconsistent state, because it would have
6835  * incremented some queue counters for an rq destined to
6836  * transformation, without any chance to correctly lower these
6837  * counters back. In contrast, no transformation can still happen for
6838  * rq after rq has been inserted or merged. So, it is safe to execute
6839  * these preparation operations when rq is finally inserted or merged.
6840  */
6841 static struct bfq_queue *bfq_init_rq(struct request *rq)
6842 {
6843 	struct request_queue *q = rq->q;
6844 	struct bio *bio = rq->bio;
6845 	struct bfq_data *bfqd = q->elevator->elevator_data;
6846 	struct bfq_io_cq *bic;
6847 	const int is_sync = rq_is_sync(rq);
6848 	struct bfq_queue *bfqq;
6849 	bool new_queue = false;
6850 	bool bfqq_already_existing = false, split = false;
6851 	unsigned int a_idx = bfq_actuator_index(bfqd, bio);
6852 
6853 	if (unlikely(!rq->elv.icq))
6854 		return NULL;
6855 
6856 	/*
6857 	 * Assuming that RQ_BFQQ(rq) is set only if everything is set
6858 	 * for this rq. This holds true, because this function is
6859 	 * invoked only for insertion or merging, and, after such
6860 	 * events, a request cannot be manipulated any longer before
6861 	 * being removed from bfq.
6862 	 */
6863 	if (RQ_BFQQ(rq))
6864 		return RQ_BFQQ(rq);
6865 
6866 	bic = icq_to_bic(rq->elv.icq);
6867 
6868 	bfq_check_ioprio_change(bic, bio);
6869 
6870 	bfq_bic_update_cgroup(bic, bio);
6871 
6872 	bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio, false, is_sync,
6873 					 &new_queue);
6874 
6875 	if (likely(!new_queue)) {
6876 		/* If the queue was seeky for too long, break it apart. */
6877 		if (bfq_bfqq_coop(bfqq) && bfq_bfqq_split_coop(bfqq) &&
6878 			!bic->bfqq_data[a_idx].stably_merged) {
6879 			struct bfq_queue *old_bfqq = bfqq;
6880 
6881 			/* Update bic before losing reference to bfqq */
6882 			if (bfq_bfqq_in_large_burst(bfqq))
6883 				bic->bfqq_data[a_idx].saved_in_large_burst =
6884 					true;
6885 
6886 			bfqq = bfq_split_bfqq(bic, bfqq);
6887 			split = true;
6888 
6889 			if (!bfqq) {
6890 				bfqq = bfq_get_bfqq_handle_split(bfqd, bic, bio,
6891 								 true, is_sync,
6892 								 NULL);
6893 				if (unlikely(bfqq == &bfqd->oom_bfqq))
6894 					bfqq_already_existing = true;
6895 			} else
6896 				bfqq_already_existing = true;
6897 
6898 			if (!bfqq_already_existing) {
6899 				bfqq->waker_bfqq = old_bfqq->waker_bfqq;
6900 				bfqq->tentative_waker_bfqq = NULL;
6901 
6902 				/*
6903 				 * If the waker queue disappears, then
6904 				 * new_bfqq->waker_bfqq must be
6905 				 * reset. So insert new_bfqq into the
6906 				 * woken_list of the waker. See
6907 				 * bfq_check_waker for details.
6908 				 */
6909 				if (bfqq->waker_bfqq)
6910 					hlist_add_head(&bfqq->woken_list_node,
6911 						       &bfqq->waker_bfqq->woken_list);
6912 			}
6913 		}
6914 	}
6915 
6916 	bfqq_request_allocated(bfqq);
6917 	bfqq->ref++;
6918 	bic->requests++;
6919 	bfq_log_bfqq(bfqd, bfqq, "get_request %p: bfqq %p, %d",
6920 		     rq, bfqq, bfqq->ref);
6921 
6922 	rq->elv.priv[0] = bic;
6923 	rq->elv.priv[1] = bfqq;
6924 
6925 	/*
6926 	 * If a bfq_queue has only one process reference, it is owned
6927 	 * by only this bic: we can then set bfqq->bic = bic. in
6928 	 * addition, if the queue has also just been split, we have to
6929 	 * resume its state.
6930 	 */
6931 	if (likely(bfqq != &bfqd->oom_bfqq) && bfqq_process_refs(bfqq) == 1) {
6932 		bfqq->bic = bic;
6933 		if (split) {
6934 			/*
6935 			 * The queue has just been split from a shared
6936 			 * queue: restore the idle window and the
6937 			 * possible weight raising period.
6938 			 */
6939 			bfq_bfqq_resume_state(bfqq, bfqd, bic,
6940 					      bfqq_already_existing);
6941 		}
6942 	}
6943 
6944 	/*
6945 	 * Consider bfqq as possibly belonging to a burst of newly
6946 	 * created queues only if:
6947 	 * 1) A burst is actually happening (bfqd->burst_size > 0)
6948 	 * or
6949 	 * 2) There is no other active queue. In fact, if, in
6950 	 *    contrast, there are active queues not belonging to the
6951 	 *    possible burst bfqq may belong to, then there is no gain
6952 	 *    in considering bfqq as belonging to a burst, and
6953 	 *    therefore in not weight-raising bfqq. See comments on
6954 	 *    bfq_handle_burst().
6955 	 *
6956 	 * This filtering also helps eliminating false positives,
6957 	 * occurring when bfqq does not belong to an actual large
6958 	 * burst, but some background task (e.g., a service) happens
6959 	 * to trigger the creation of new queues very close to when
6960 	 * bfqq and its possible companion queues are created. See
6961 	 * comments on bfq_handle_burst() for further details also on
6962 	 * this issue.
6963 	 */
6964 	if (unlikely(bfq_bfqq_just_created(bfqq) &&
6965 		     (bfqd->burst_size > 0 ||
6966 		      bfq_tot_busy_queues(bfqd) == 0)))
6967 		bfq_handle_burst(bfqd, bfqq);
6968 
6969 	return bfqq;
6970 }
6971 
6972 static void
6973 bfq_idle_slice_timer_body(struct bfq_data *bfqd, struct bfq_queue *bfqq)
6974 {
6975 	enum bfqq_expiration reason;
6976 	unsigned long flags;
6977 
6978 	spin_lock_irqsave(&bfqd->lock, flags);
6979 
6980 	/*
6981 	 * Considering that bfqq may be in race, we should firstly check
6982 	 * whether bfqq is in service before doing something on it. If
6983 	 * the bfqq in race is not in service, it has already been expired
6984 	 * through __bfq_bfqq_expire func and its wait_request flags has
6985 	 * been cleared in __bfq_bfqd_reset_in_service func.
6986 	 */
6987 	if (bfqq != bfqd->in_service_queue) {
6988 		spin_unlock_irqrestore(&bfqd->lock, flags);
6989 		return;
6990 	}
6991 
6992 	bfq_clear_bfqq_wait_request(bfqq);
6993 
6994 	if (bfq_bfqq_budget_timeout(bfqq))
6995 		/*
6996 		 * Also here the queue can be safely expired
6997 		 * for budget timeout without wasting
6998 		 * guarantees
6999 		 */
7000 		reason = BFQQE_BUDGET_TIMEOUT;
7001 	else if (bfqq->queued[0] == 0 && bfqq->queued[1] == 0)
7002 		/*
7003 		 * The queue may not be empty upon timer expiration,
7004 		 * because we may not disable the timer when the
7005 		 * first request of the in-service queue arrives
7006 		 * during disk idling.
7007 		 */
7008 		reason = BFQQE_TOO_IDLE;
7009 	else
7010 		goto schedule_dispatch;
7011 
7012 	bfq_bfqq_expire(bfqd, bfqq, true, reason);
7013 
7014 schedule_dispatch:
7015 	bfq_schedule_dispatch(bfqd);
7016 	spin_unlock_irqrestore(&bfqd->lock, flags);
7017 }
7018 
7019 /*
7020  * Handler of the expiration of the timer running if the in-service queue
7021  * is idling inside its time slice.
7022  */
7023 static enum hrtimer_restart bfq_idle_slice_timer(struct hrtimer *timer)
7024 {
7025 	struct bfq_data *bfqd = container_of(timer, struct bfq_data,
7026 					     idle_slice_timer);
7027 	struct bfq_queue *bfqq = bfqd->in_service_queue;
7028 
7029 	/*
7030 	 * Theoretical race here: the in-service queue can be NULL or
7031 	 * different from the queue that was idling if a new request
7032 	 * arrives for the current queue and there is a full dispatch
7033 	 * cycle that changes the in-service queue.  This can hardly
7034 	 * happen, but in the worst case we just expire a queue too
7035 	 * early.
7036 	 */
7037 	if (bfqq)
7038 		bfq_idle_slice_timer_body(bfqd, bfqq);
7039 
7040 	return HRTIMER_NORESTART;
7041 }
7042 
7043 static void __bfq_put_async_bfqq(struct bfq_data *bfqd,
7044 				 struct bfq_queue **bfqq_ptr)
7045 {
7046 	struct bfq_queue *bfqq = *bfqq_ptr;
7047 
7048 	bfq_log(bfqd, "put_async_bfqq: %p", bfqq);
7049 	if (bfqq) {
7050 		bfq_bfqq_move(bfqd, bfqq, bfqd->root_group);
7051 
7052 		bfq_log_bfqq(bfqd, bfqq, "put_async_bfqq: putting %p, %d",
7053 			     bfqq, bfqq->ref);
7054 		bfq_put_queue(bfqq);
7055 		*bfqq_ptr = NULL;
7056 	}
7057 }
7058 
7059 /*
7060  * Release all the bfqg references to its async queues.  If we are
7061  * deallocating the group these queues may still contain requests, so
7062  * we reparent them to the root cgroup (i.e., the only one that will
7063  * exist for sure until all the requests on a device are gone).
7064  */
7065 void bfq_put_async_queues(struct bfq_data *bfqd, struct bfq_group *bfqg)
7066 {
7067 	int i, j, k;
7068 
7069 	for (k = 0; k < bfqd->num_actuators; k++) {
7070 		for (i = 0; i < 2; i++)
7071 			for (j = 0; j < IOPRIO_NR_LEVELS; j++)
7072 				__bfq_put_async_bfqq(bfqd, &bfqg->async_bfqq[i][j][k]);
7073 
7074 		__bfq_put_async_bfqq(bfqd, &bfqg->async_idle_bfqq[k]);
7075 	}
7076 }
7077 
7078 /*
7079  * See the comments on bfq_limit_depth for the purpose of
7080  * the depths set in the function. Return minimum shallow depth we'll use.
7081  */
7082 static void bfq_update_depths(struct bfq_data *bfqd, struct sbitmap_queue *bt)
7083 {
7084 	unsigned int depth = 1U << bt->sb.shift;
7085 
7086 	bfqd->full_depth_shift = bt->sb.shift;
7087 	/*
7088 	 * In-word depths if no bfq_queue is being weight-raised:
7089 	 * leaving 25% of tags only for sync reads.
7090 	 *
7091 	 * In next formulas, right-shift the value
7092 	 * (1U<<bt->sb.shift), instead of computing directly
7093 	 * (1U<<(bt->sb.shift - something)), to be robust against
7094 	 * any possible value of bt->sb.shift, without having to
7095 	 * limit 'something'.
7096 	 */
7097 	/* no more than 50% of tags for async I/O */
7098 	bfqd->word_depths[0][0] = max(depth >> 1, 1U);
7099 	/*
7100 	 * no more than 75% of tags for sync writes (25% extra tags
7101 	 * w.r.t. async I/O, to prevent async I/O from starving sync
7102 	 * writes)
7103 	 */
7104 	bfqd->word_depths[0][1] = max((depth * 3) >> 2, 1U);
7105 
7106 	/*
7107 	 * In-word depths in case some bfq_queue is being weight-
7108 	 * raised: leaving ~63% of tags for sync reads. This is the
7109 	 * highest percentage for which, in our tests, application
7110 	 * start-up times didn't suffer from any regression due to tag
7111 	 * shortage.
7112 	 */
7113 	/* no more than ~18% of tags for async I/O */
7114 	bfqd->word_depths[1][0] = max((depth * 3) >> 4, 1U);
7115 	/* no more than ~37% of tags for sync writes (~20% extra tags) */
7116 	bfqd->word_depths[1][1] = max((depth * 6) >> 4, 1U);
7117 }
7118 
7119 static void bfq_depth_updated(struct blk_mq_hw_ctx *hctx)
7120 {
7121 	struct bfq_data *bfqd = hctx->queue->elevator->elevator_data;
7122 	struct blk_mq_tags *tags = hctx->sched_tags;
7123 
7124 	bfq_update_depths(bfqd, &tags->bitmap_tags);
7125 	sbitmap_queue_min_shallow_depth(&tags->bitmap_tags, 1);
7126 }
7127 
7128 static int bfq_init_hctx(struct blk_mq_hw_ctx *hctx, unsigned int index)
7129 {
7130 	bfq_depth_updated(hctx);
7131 	return 0;
7132 }
7133 
7134 static void bfq_exit_queue(struct elevator_queue *e)
7135 {
7136 	struct bfq_data *bfqd = e->elevator_data;
7137 	struct bfq_queue *bfqq, *n;
7138 
7139 	hrtimer_cancel(&bfqd->idle_slice_timer);
7140 
7141 	spin_lock_irq(&bfqd->lock);
7142 	list_for_each_entry_safe(bfqq, n, &bfqd->idle_list, bfqq_list)
7143 		bfq_deactivate_bfqq(bfqd, bfqq, false, false);
7144 	spin_unlock_irq(&bfqd->lock);
7145 
7146 	hrtimer_cancel(&bfqd->idle_slice_timer);
7147 
7148 	/* release oom-queue reference to root group */
7149 	bfqg_and_blkg_put(bfqd->root_group);
7150 
7151 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7152 	blkcg_deactivate_policy(bfqd->queue->disk, &blkcg_policy_bfq);
7153 #else
7154 	spin_lock_irq(&bfqd->lock);
7155 	bfq_put_async_queues(bfqd, bfqd->root_group);
7156 	kfree(bfqd->root_group);
7157 	spin_unlock_irq(&bfqd->lock);
7158 #endif
7159 
7160 	blk_stat_disable_accounting(bfqd->queue);
7161 	clear_bit(ELEVATOR_FLAG_DISABLE_WBT, &e->flags);
7162 	wbt_enable_default(bfqd->queue->disk);
7163 
7164 	kfree(bfqd);
7165 }
7166 
7167 static void bfq_init_root_group(struct bfq_group *root_group,
7168 				struct bfq_data *bfqd)
7169 {
7170 	int i;
7171 
7172 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7173 	root_group->entity.parent = NULL;
7174 	root_group->my_entity = NULL;
7175 	root_group->bfqd = bfqd;
7176 #endif
7177 	root_group->rq_pos_tree = RB_ROOT;
7178 	for (i = 0; i < BFQ_IOPRIO_CLASSES; i++)
7179 		root_group->sched_data.service_tree[i] = BFQ_SERVICE_TREE_INIT;
7180 	root_group->sched_data.bfq_class_idle_last_service = jiffies;
7181 }
7182 
7183 static int bfq_init_queue(struct request_queue *q, struct elevator_type *e)
7184 {
7185 	struct bfq_data *bfqd;
7186 	struct elevator_queue *eq;
7187 	unsigned int i;
7188 	struct blk_independent_access_ranges *ia_ranges = q->disk->ia_ranges;
7189 
7190 	eq = elevator_alloc(q, e);
7191 	if (!eq)
7192 		return -ENOMEM;
7193 
7194 	bfqd = kzalloc_node(sizeof(*bfqd), GFP_KERNEL, q->node);
7195 	if (!bfqd) {
7196 		kobject_put(&eq->kobj);
7197 		return -ENOMEM;
7198 	}
7199 	eq->elevator_data = bfqd;
7200 
7201 	spin_lock_irq(&q->queue_lock);
7202 	q->elevator = eq;
7203 	spin_unlock_irq(&q->queue_lock);
7204 
7205 	/*
7206 	 * Our fallback bfqq if bfq_find_alloc_queue() runs into OOM issues.
7207 	 * Grab a permanent reference to it, so that the normal code flow
7208 	 * will not attempt to free it.
7209 	 * Set zero as actuator index: we will pretend that
7210 	 * all I/O requests are for the same actuator.
7211 	 */
7212 	bfq_init_bfqq(bfqd, &bfqd->oom_bfqq, NULL, 1, 0, 0);
7213 	bfqd->oom_bfqq.ref++;
7214 	bfqd->oom_bfqq.new_ioprio = BFQ_DEFAULT_QUEUE_IOPRIO;
7215 	bfqd->oom_bfqq.new_ioprio_class = IOPRIO_CLASS_BE;
7216 	bfqd->oom_bfqq.entity.new_weight =
7217 		bfq_ioprio_to_weight(bfqd->oom_bfqq.new_ioprio);
7218 
7219 	/* oom_bfqq does not participate to bursts */
7220 	bfq_clear_bfqq_just_created(&bfqd->oom_bfqq);
7221 
7222 	/*
7223 	 * Trigger weight initialization, according to ioprio, at the
7224 	 * oom_bfqq's first activation. The oom_bfqq's ioprio and ioprio
7225 	 * class won't be changed any more.
7226 	 */
7227 	bfqd->oom_bfqq.entity.prio_changed = 1;
7228 
7229 	bfqd->queue = q;
7230 
7231 	bfqd->num_actuators = 1;
7232 	/*
7233 	 * If the disk supports multiple actuators, copy independent
7234 	 * access ranges from the request queue structure.
7235 	 */
7236 	spin_lock_irq(&q->queue_lock);
7237 	if (ia_ranges) {
7238 		/*
7239 		 * Check if the disk ia_ranges size exceeds the current bfq
7240 		 * actuator limit.
7241 		 */
7242 		if (ia_ranges->nr_ia_ranges > BFQ_MAX_ACTUATORS) {
7243 			pr_crit("nr_ia_ranges higher than act limit: iars=%d, max=%d.\n",
7244 				ia_ranges->nr_ia_ranges, BFQ_MAX_ACTUATORS);
7245 			pr_crit("Falling back to single actuator mode.\n");
7246 		} else {
7247 			bfqd->num_actuators = ia_ranges->nr_ia_ranges;
7248 
7249 			for (i = 0; i < bfqd->num_actuators; i++) {
7250 				bfqd->sector[i] = ia_ranges->ia_range[i].sector;
7251 				bfqd->nr_sectors[i] =
7252 					ia_ranges->ia_range[i].nr_sectors;
7253 			}
7254 		}
7255 	}
7256 
7257 	/* Otherwise use single-actuator dev info */
7258 	if (bfqd->num_actuators == 1) {
7259 		bfqd->sector[0] = 0;
7260 		bfqd->nr_sectors[0] = get_capacity(q->disk);
7261 	}
7262 	spin_unlock_irq(&q->queue_lock);
7263 
7264 	INIT_LIST_HEAD(&bfqd->dispatch);
7265 
7266 	hrtimer_init(&bfqd->idle_slice_timer, CLOCK_MONOTONIC,
7267 		     HRTIMER_MODE_REL);
7268 	bfqd->idle_slice_timer.function = bfq_idle_slice_timer;
7269 
7270 	bfqd->queue_weights_tree = RB_ROOT_CACHED;
7271 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7272 	bfqd->num_groups_with_pending_reqs = 0;
7273 #endif
7274 
7275 	INIT_LIST_HEAD(&bfqd->active_list[0]);
7276 	INIT_LIST_HEAD(&bfqd->active_list[1]);
7277 	INIT_LIST_HEAD(&bfqd->idle_list);
7278 	INIT_HLIST_HEAD(&bfqd->burst_list);
7279 
7280 	bfqd->hw_tag = -1;
7281 	bfqd->nonrot_with_queueing = blk_queue_nonrot(bfqd->queue);
7282 
7283 	bfqd->bfq_max_budget = bfq_default_max_budget;
7284 
7285 	bfqd->bfq_fifo_expire[0] = bfq_fifo_expire[0];
7286 	bfqd->bfq_fifo_expire[1] = bfq_fifo_expire[1];
7287 	bfqd->bfq_back_max = bfq_back_max;
7288 	bfqd->bfq_back_penalty = bfq_back_penalty;
7289 	bfqd->bfq_slice_idle = bfq_slice_idle;
7290 	bfqd->bfq_timeout = bfq_timeout;
7291 
7292 	bfqd->bfq_large_burst_thresh = 8;
7293 	bfqd->bfq_burst_interval = msecs_to_jiffies(180);
7294 
7295 	bfqd->low_latency = true;
7296 
7297 	/*
7298 	 * Trade-off between responsiveness and fairness.
7299 	 */
7300 	bfqd->bfq_wr_coeff = 30;
7301 	bfqd->bfq_wr_rt_max_time = msecs_to_jiffies(300);
7302 	bfqd->bfq_wr_min_idle_time = msecs_to_jiffies(2000);
7303 	bfqd->bfq_wr_min_inter_arr_async = msecs_to_jiffies(500);
7304 	bfqd->bfq_wr_max_softrt_rate = 7000; /*
7305 					      * Approximate rate required
7306 					      * to playback or record a
7307 					      * high-definition compressed
7308 					      * video.
7309 					      */
7310 	bfqd->wr_busy_queues = 0;
7311 
7312 	/*
7313 	 * Begin by assuming, optimistically, that the device peak
7314 	 * rate is equal to 2/3 of the highest reference rate.
7315 	 */
7316 	bfqd->rate_dur_prod = ref_rate[blk_queue_nonrot(bfqd->queue)] *
7317 		ref_wr_duration[blk_queue_nonrot(bfqd->queue)];
7318 	bfqd->peak_rate = ref_rate[blk_queue_nonrot(bfqd->queue)] * 2 / 3;
7319 
7320 	/* see comments on the definition of next field inside bfq_data */
7321 	bfqd->actuator_load_threshold = 4;
7322 
7323 	spin_lock_init(&bfqd->lock);
7324 
7325 	/*
7326 	 * The invocation of the next bfq_create_group_hierarchy
7327 	 * function is the head of a chain of function calls
7328 	 * (bfq_create_group_hierarchy->blkcg_activate_policy->
7329 	 * blk_mq_freeze_queue) that may lead to the invocation of the
7330 	 * has_work hook function. For this reason,
7331 	 * bfq_create_group_hierarchy is invoked only after all
7332 	 * scheduler data has been initialized, apart from the fields
7333 	 * that can be initialized only after invoking
7334 	 * bfq_create_group_hierarchy. This, in particular, enables
7335 	 * has_work to correctly return false. Of course, to avoid
7336 	 * other inconsistencies, the blk-mq stack must then refrain
7337 	 * from invoking further scheduler hooks before this init
7338 	 * function is finished.
7339 	 */
7340 	bfqd->root_group = bfq_create_group_hierarchy(bfqd, q->node);
7341 	if (!bfqd->root_group)
7342 		goto out_free;
7343 	bfq_init_root_group(bfqd->root_group, bfqd);
7344 	bfq_init_entity(&bfqd->oom_bfqq.entity, bfqd->root_group);
7345 
7346 	/* We dispatch from request queue wide instead of hw queue */
7347 	blk_queue_flag_set(QUEUE_FLAG_SQ_SCHED, q);
7348 
7349 	set_bit(ELEVATOR_FLAG_DISABLE_WBT, &eq->flags);
7350 	wbt_disable_default(q->disk);
7351 	blk_stat_enable_accounting(q);
7352 
7353 	return 0;
7354 
7355 out_free:
7356 	kfree(bfqd);
7357 	kobject_put(&eq->kobj);
7358 	return -ENOMEM;
7359 }
7360 
7361 static void bfq_slab_kill(void)
7362 {
7363 	kmem_cache_destroy(bfq_pool);
7364 }
7365 
7366 static int __init bfq_slab_setup(void)
7367 {
7368 	bfq_pool = KMEM_CACHE(bfq_queue, 0);
7369 	if (!bfq_pool)
7370 		return -ENOMEM;
7371 	return 0;
7372 }
7373 
7374 static ssize_t bfq_var_show(unsigned int var, char *page)
7375 {
7376 	return sprintf(page, "%u\n", var);
7377 }
7378 
7379 static int bfq_var_store(unsigned long *var, const char *page)
7380 {
7381 	unsigned long new_val;
7382 	int ret = kstrtoul(page, 10, &new_val);
7383 
7384 	if (ret)
7385 		return ret;
7386 	*var = new_val;
7387 	return 0;
7388 }
7389 
7390 #define SHOW_FUNCTION(__FUNC, __VAR, __CONV)				\
7391 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
7392 {									\
7393 	struct bfq_data *bfqd = e->elevator_data;			\
7394 	u64 __data = __VAR;						\
7395 	if (__CONV == 1)						\
7396 		__data = jiffies_to_msecs(__data);			\
7397 	else if (__CONV == 2)						\
7398 		__data = div_u64(__data, NSEC_PER_MSEC);		\
7399 	return bfq_var_show(__data, (page));				\
7400 }
7401 SHOW_FUNCTION(bfq_fifo_expire_sync_show, bfqd->bfq_fifo_expire[1], 2);
7402 SHOW_FUNCTION(bfq_fifo_expire_async_show, bfqd->bfq_fifo_expire[0], 2);
7403 SHOW_FUNCTION(bfq_back_seek_max_show, bfqd->bfq_back_max, 0);
7404 SHOW_FUNCTION(bfq_back_seek_penalty_show, bfqd->bfq_back_penalty, 0);
7405 SHOW_FUNCTION(bfq_slice_idle_show, bfqd->bfq_slice_idle, 2);
7406 SHOW_FUNCTION(bfq_max_budget_show, bfqd->bfq_user_max_budget, 0);
7407 SHOW_FUNCTION(bfq_timeout_sync_show, bfqd->bfq_timeout, 1);
7408 SHOW_FUNCTION(bfq_strict_guarantees_show, bfqd->strict_guarantees, 0);
7409 SHOW_FUNCTION(bfq_low_latency_show, bfqd->low_latency, 0);
7410 #undef SHOW_FUNCTION
7411 
7412 #define USEC_SHOW_FUNCTION(__FUNC, __VAR)				\
7413 static ssize_t __FUNC(struct elevator_queue *e, char *page)		\
7414 {									\
7415 	struct bfq_data *bfqd = e->elevator_data;			\
7416 	u64 __data = __VAR;						\
7417 	__data = div_u64(__data, NSEC_PER_USEC);			\
7418 	return bfq_var_show(__data, (page));				\
7419 }
7420 USEC_SHOW_FUNCTION(bfq_slice_idle_us_show, bfqd->bfq_slice_idle);
7421 #undef USEC_SHOW_FUNCTION
7422 
7423 #define STORE_FUNCTION(__FUNC, __PTR, MIN, MAX, __CONV)			\
7424 static ssize_t								\
7425 __FUNC(struct elevator_queue *e, const char *page, size_t count)	\
7426 {									\
7427 	struct bfq_data *bfqd = e->elevator_data;			\
7428 	unsigned long __data, __min = (MIN), __max = (MAX);		\
7429 	int ret;							\
7430 									\
7431 	ret = bfq_var_store(&__data, (page));				\
7432 	if (ret)							\
7433 		return ret;						\
7434 	if (__data < __min)						\
7435 		__data = __min;						\
7436 	else if (__data > __max)					\
7437 		__data = __max;						\
7438 	if (__CONV == 1)						\
7439 		*(__PTR) = msecs_to_jiffies(__data);			\
7440 	else if (__CONV == 2)						\
7441 		*(__PTR) = (u64)__data * NSEC_PER_MSEC;			\
7442 	else								\
7443 		*(__PTR) = __data;					\
7444 	return count;							\
7445 }
7446 STORE_FUNCTION(bfq_fifo_expire_sync_store, &bfqd->bfq_fifo_expire[1], 1,
7447 		INT_MAX, 2);
7448 STORE_FUNCTION(bfq_fifo_expire_async_store, &bfqd->bfq_fifo_expire[0], 1,
7449 		INT_MAX, 2);
7450 STORE_FUNCTION(bfq_back_seek_max_store, &bfqd->bfq_back_max, 0, INT_MAX, 0);
7451 STORE_FUNCTION(bfq_back_seek_penalty_store, &bfqd->bfq_back_penalty, 1,
7452 		INT_MAX, 0);
7453 STORE_FUNCTION(bfq_slice_idle_store, &bfqd->bfq_slice_idle, 0, INT_MAX, 2);
7454 #undef STORE_FUNCTION
7455 
7456 #define USEC_STORE_FUNCTION(__FUNC, __PTR, MIN, MAX)			\
7457 static ssize_t __FUNC(struct elevator_queue *e, const char *page, size_t count)\
7458 {									\
7459 	struct bfq_data *bfqd = e->elevator_data;			\
7460 	unsigned long __data, __min = (MIN), __max = (MAX);		\
7461 	int ret;							\
7462 									\
7463 	ret = bfq_var_store(&__data, (page));				\
7464 	if (ret)							\
7465 		return ret;						\
7466 	if (__data < __min)						\
7467 		__data = __min;						\
7468 	else if (__data > __max)					\
7469 		__data = __max;						\
7470 	*(__PTR) = (u64)__data * NSEC_PER_USEC;				\
7471 	return count;							\
7472 }
7473 USEC_STORE_FUNCTION(bfq_slice_idle_us_store, &bfqd->bfq_slice_idle, 0,
7474 		    UINT_MAX);
7475 #undef USEC_STORE_FUNCTION
7476 
7477 static ssize_t bfq_max_budget_store(struct elevator_queue *e,
7478 				    const char *page, size_t count)
7479 {
7480 	struct bfq_data *bfqd = e->elevator_data;
7481 	unsigned long __data;
7482 	int ret;
7483 
7484 	ret = bfq_var_store(&__data, (page));
7485 	if (ret)
7486 		return ret;
7487 
7488 	if (__data == 0)
7489 		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7490 	else {
7491 		if (__data > INT_MAX)
7492 			__data = INT_MAX;
7493 		bfqd->bfq_max_budget = __data;
7494 	}
7495 
7496 	bfqd->bfq_user_max_budget = __data;
7497 
7498 	return count;
7499 }
7500 
7501 /*
7502  * Leaving this name to preserve name compatibility with cfq
7503  * parameters, but this timeout is used for both sync and async.
7504  */
7505 static ssize_t bfq_timeout_sync_store(struct elevator_queue *e,
7506 				      const char *page, size_t count)
7507 {
7508 	struct bfq_data *bfqd = e->elevator_data;
7509 	unsigned long __data;
7510 	int ret;
7511 
7512 	ret = bfq_var_store(&__data, (page));
7513 	if (ret)
7514 		return ret;
7515 
7516 	if (__data < 1)
7517 		__data = 1;
7518 	else if (__data > INT_MAX)
7519 		__data = INT_MAX;
7520 
7521 	bfqd->bfq_timeout = msecs_to_jiffies(__data);
7522 	if (bfqd->bfq_user_max_budget == 0)
7523 		bfqd->bfq_max_budget = bfq_calc_max_budget(bfqd);
7524 
7525 	return count;
7526 }
7527 
7528 static ssize_t bfq_strict_guarantees_store(struct elevator_queue *e,
7529 				     const char *page, size_t count)
7530 {
7531 	struct bfq_data *bfqd = e->elevator_data;
7532 	unsigned long __data;
7533 	int ret;
7534 
7535 	ret = bfq_var_store(&__data, (page));
7536 	if (ret)
7537 		return ret;
7538 
7539 	if (__data > 1)
7540 		__data = 1;
7541 	if (!bfqd->strict_guarantees && __data == 1
7542 	    && bfqd->bfq_slice_idle < 8 * NSEC_PER_MSEC)
7543 		bfqd->bfq_slice_idle = 8 * NSEC_PER_MSEC;
7544 
7545 	bfqd->strict_guarantees = __data;
7546 
7547 	return count;
7548 }
7549 
7550 static ssize_t bfq_low_latency_store(struct elevator_queue *e,
7551 				     const char *page, size_t count)
7552 {
7553 	struct bfq_data *bfqd = e->elevator_data;
7554 	unsigned long __data;
7555 	int ret;
7556 
7557 	ret = bfq_var_store(&__data, (page));
7558 	if (ret)
7559 		return ret;
7560 
7561 	if (__data > 1)
7562 		__data = 1;
7563 	if (__data == 0 && bfqd->low_latency != 0)
7564 		bfq_end_wr(bfqd);
7565 	bfqd->low_latency = __data;
7566 
7567 	return count;
7568 }
7569 
7570 #define BFQ_ATTR(name) \
7571 	__ATTR(name, 0644, bfq_##name##_show, bfq_##name##_store)
7572 
7573 static struct elv_fs_entry bfq_attrs[] = {
7574 	BFQ_ATTR(fifo_expire_sync),
7575 	BFQ_ATTR(fifo_expire_async),
7576 	BFQ_ATTR(back_seek_max),
7577 	BFQ_ATTR(back_seek_penalty),
7578 	BFQ_ATTR(slice_idle),
7579 	BFQ_ATTR(slice_idle_us),
7580 	BFQ_ATTR(max_budget),
7581 	BFQ_ATTR(timeout_sync),
7582 	BFQ_ATTR(strict_guarantees),
7583 	BFQ_ATTR(low_latency),
7584 	__ATTR_NULL
7585 };
7586 
7587 static struct elevator_type iosched_bfq_mq = {
7588 	.ops = {
7589 		.limit_depth		= bfq_limit_depth,
7590 		.prepare_request	= bfq_prepare_request,
7591 		.requeue_request        = bfq_finish_requeue_request,
7592 		.finish_request		= bfq_finish_request,
7593 		.exit_icq		= bfq_exit_icq,
7594 		.insert_requests	= bfq_insert_requests,
7595 		.dispatch_request	= bfq_dispatch_request,
7596 		.next_request		= elv_rb_latter_request,
7597 		.former_request		= elv_rb_former_request,
7598 		.allow_merge		= bfq_allow_bio_merge,
7599 		.bio_merge		= bfq_bio_merge,
7600 		.request_merge		= bfq_request_merge,
7601 		.requests_merged	= bfq_requests_merged,
7602 		.request_merged		= bfq_request_merged,
7603 		.has_work		= bfq_has_work,
7604 		.depth_updated		= bfq_depth_updated,
7605 		.init_hctx		= bfq_init_hctx,
7606 		.init_sched		= bfq_init_queue,
7607 		.exit_sched		= bfq_exit_queue,
7608 	},
7609 
7610 	.icq_size =		sizeof(struct bfq_io_cq),
7611 	.icq_align =		__alignof__(struct bfq_io_cq),
7612 	.elevator_attrs =	bfq_attrs,
7613 	.elevator_name =	"bfq",
7614 	.elevator_owner =	THIS_MODULE,
7615 };
7616 MODULE_ALIAS("bfq-iosched");
7617 
7618 static int __init bfq_init(void)
7619 {
7620 	int ret;
7621 
7622 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7623 	ret = blkcg_policy_register(&blkcg_policy_bfq);
7624 	if (ret)
7625 		return ret;
7626 #endif
7627 
7628 	ret = -ENOMEM;
7629 	if (bfq_slab_setup())
7630 		goto err_pol_unreg;
7631 
7632 	/*
7633 	 * Times to load large popular applications for the typical
7634 	 * systems installed on the reference devices (see the
7635 	 * comments before the definition of the next
7636 	 * array). Actually, we use slightly lower values, as the
7637 	 * estimated peak rate tends to be smaller than the actual
7638 	 * peak rate.  The reason for this last fact is that estimates
7639 	 * are computed over much shorter time intervals than the long
7640 	 * intervals typically used for benchmarking. Why? First, to
7641 	 * adapt more quickly to variations. Second, because an I/O
7642 	 * scheduler cannot rely on a peak-rate-evaluation workload to
7643 	 * be run for a long time.
7644 	 */
7645 	ref_wr_duration[0] = msecs_to_jiffies(7000); /* actually 8 sec */
7646 	ref_wr_duration[1] = msecs_to_jiffies(2500); /* actually 3 sec */
7647 
7648 	ret = elv_register(&iosched_bfq_mq);
7649 	if (ret)
7650 		goto slab_kill;
7651 
7652 	return 0;
7653 
7654 slab_kill:
7655 	bfq_slab_kill();
7656 err_pol_unreg:
7657 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7658 	blkcg_policy_unregister(&blkcg_policy_bfq);
7659 #endif
7660 	return ret;
7661 }
7662 
7663 static void __exit bfq_exit(void)
7664 {
7665 	elv_unregister(&iosched_bfq_mq);
7666 #ifdef CONFIG_BFQ_GROUP_IOSCHED
7667 	blkcg_policy_unregister(&blkcg_policy_bfq);
7668 #endif
7669 	bfq_slab_kill();
7670 }
7671 
7672 module_init(bfq_init);
7673 module_exit(bfq_exit);
7674 
7675 MODULE_AUTHOR("Paolo Valente");
7676 MODULE_LICENSE("GPL");
7677 MODULE_DESCRIPTION("MQ Budget Fair Queueing I/O Scheduler");
7678